[2026-08-07] ingest | Prompting Guide techniques index (backlog cleanup)

Source

raw/external/www-promptingguide-ai-techniques-f9bd0064.md — Prompting Guide techniques landing/index page (English snapshot, same catalog already cited by individual technique articles)

Route validation

scripts/route-classifier.ps1 -Query "prompt engineering techniques from promptingguide.ai" → ambiguous (0.603), main 6.038 / prompts 9.157, top match Prompt Engineering Guide (promptingguide.ai). Override → prompt wiki — user explicit decision; content is a prompt techniques index page.

Discarded (not ingested — no wiki coverage needed)

raw/external/developers-openai-com-9dcdba4f.md (cc1760193b07a0c7) — OpenAI developers homepage/API docs landing; too generic, no unique evidence for prompt wiki coverage.
raw/feeds/https-news-ycombinator-com-item-id-49162086-2dfa28a790db9021.md (f5c2bff46dae906b) — HN off-topic “Windows XP Itanium” post; outside both wikis’ scope.

Verification

wikilint: OK (pending final run below)

[2026-08-07] query | Polimento de prompt de pesquisa (Superpowers × Matt Pocock)

Routing: scripts/route-classifier.ps1 -Query "apply prompt wiki knowledge to polish a deep research prompt comparing Superpowers vs Matt Pocock workflow" → ambiguous (0.502), main 14.02 / prompts 13.93. Override manual → prompt wiki: a tarefa é explicitamente aplicar conhecimento de prompt engineering para polir um prompt.
Retrieval (prompt wiki): documented-research-process, how-to-structure-a-system-prompt, constraint-based-prompting, communication-style-spectrum, gemini-deep-research-prompt, two-tier-search-architecture. Found: 10 | Analyzed: 6 core.
Deliverable: prompt de pesquisa polido em PT-BR aplicando as 5 seções universais, Task State Machine (TDP/MIG/CAS/RG/IR), constraint-based prompting, communication style explain-before-do-concise, e Memory Block. Atingiu o prompt wiki’s [[gemini-deep-research-prompt]] para a estrutura de pesquisa.
Gate: scripts/query-check.ps1 -Scope prompts → query-check: OK (0 failures; 10 retrieved slugs, all in-route).
Gains: prompt de pesquisa reutilizável e rigoroso com citação obrigatória e separação fato/opinião. Losses: nenhuma. Nenhuma página criada/alterada (consulta read-only).

[2026-08-07] query | Craft prompt to analyze and prioritize a 35-skill package

Routing: scripts/route-classifier.ps1 -Query "how to craft a prompt to analyze and prioritize skill packages" → ambiguous (0.514), main 7.34 / prompts 7.75, top match Skill Authoring Best Practices. Override manual → prompt wiki porque a tarefa é criar um prompt (domínio da prompt wiki). Classifier result noted in the answer.
Retrieval: prompt wiki concepts consulted: skill-authoring-best-practices, skills-directory-gate, constraint-based-prompting, task-state-machine, chain-of-thought-pattern, communication-style-spectrum, message-roles. Found: 7 | Analyzed: 7.
Deliverable: prompt pronto em PT-BR (papel de analista + pipeline em 5 etapas com enumerar-tudo-antes + classificação A/B/C/D por função + saída JSON obrigatória + regras de comportamento anti-hype). Dois blocos de placeholder para o founder colar os 35 skills e o objetivo do projeto.
Gate: scripts/query-check.ps1 -Scope prompts -Catalog → query-check: OK (0 failures; 10 retrieved slugs, all in-route).
Gains: prompt acionável reutilizável para avaliar qualquer pacote de skills contra qualquer projeto. Losses: nenhum. Nenhuma página criada/alterada (consulta read-only).

[2026-08-06] ingest | Tree of Thoughts technique from Prompt Engineering Guide

Source captured

raw/prompts/articles/promptingguide-pt-techniques-tot.md — conteúdo sobre Tree of Thoughts (ToT) em português
raw/external/www-promptingguide-ai-tot-fe67e69f.md — snapshot externo capturado

Route validation

scripts/route-classifier.ps1 → prompts (confidence: 0.95)
Técnica de prompting para raciocínio em árvore, cabe no prompt wiki

Pages extended

wiki/prompts/concepts/tree-of-thoughts.md — lifecycle: l0→l1, adicionadas definição completa, exemplos de “Game of 24”, avaliação de candidatos, referências

prompts-index.md já referenciava [[tree-of-thoughts]]

Verification

wikilint: OK — 149 pages

[2026-08-06] ingest | ReAct Prompting technique extension

Source

raw/external/www-promptingguide-ai-react-b247cf0b.md — English snapshot captured

Route validation

prompt wiki (confidence: 0.95)

Pages extended

wiki/prompts/concepts/react-prompting.md — lifecycle l0→l1, added How it Works, Results, Limitations, LangChain usage, Implications

Verification

wikilint: OK — 150 pages

[2026-08-06] ingest | Graphify and Capacities updates from external snapshots

Sources recaptured

raw/external/github-com-graphify-6b624867.md (recaptured)
raw/external/capacities-io-8be3cd66.md (recaptured)

Pages extended

wiki/entities/graphify.md — l1→l2, added code parsing, graph features, MCP HTTP transport, backend support, strict mode, hooks
wiki/entities/capacities.md — l0→l1, added paradigm, userbase, features, privacy, business model, notiion relationship

Verification

wikilint: OK — 265 pages (main + prompts)

[2026-08-06] ingest | nashsu/llm_wiki and claude-obsidian updates

Sources recaptured

raw/external/github-com-llm-wiki-7ec2f780.md (GitHub repo snapshot)
raw/external/github-com-claude-obsidian-4b89ed19.md (GitHub repo snapshot)

Pages extended

wiki/entities/nashsu-llm-wiki.md — l0→l1, added relationship-to-vault section
wiki/entities/claude-obsidian.md — l0→l1, added relationship-to-vault section

Verification

wikilint: OK — 265 pages

[2026-08-06] ingest | Tree of Thoughts (ToT) technique from promptingguide.ai

Source

raw/external/www-promptingguide-ai-techniques-f9bd0064.md — snapshot capturado

Route validation

prompt wiki (confidence: 0.95)

Verification

wikilint: OK — 149 pages

[2026-08-06] ingest | OpenAI Prompt Caching Guide (detailed)

Source

raw/external/developers-openai-com-prompt-caching-bc9e0f40.md — official OpenAI Prompt Caching guide (developers.openai.com)
Captured via scripts/external_links.py capture
Evidence body immutable; metadata records source URL and content hash

Route validation

scripts/route-classifier.ps1 -Query "Ingest OpenAI prompt caching guide"
Route: ambiguous (confidence: 0.622; main score: 5.3947, prompts score: 8.8712)
Manual override: prompt wiki — content is API-level prompt engineering guidance (tool-origin: openai). Overridden because the prompts score is significantly higher than the main score and top match is “Prompt Engineering Practical Guides”.

Pages updated (3 content pages, within 5-page batch cap)

wiki/prompts/concepts/prompt-caching.md — added 15 RPM/key traffic limit, GPT-5.6+ strict 1024 token minimum, and older model 1024-2048 range; promoted l1→l2
wiki/prompts/entities/gpt-5-6.md — added dedicated Prompt Caching section (implicit breakpoint changes, 1.25x write cost); promoted l1→l2
wiki/prompts/concepts/prompt-engineering-guides.md — refined OpenAI guide section with specific placement advice (static content first); promoted l1→l2

prompts-index.md — updated updated date
prompts-log.md — this entry

Guardrails applied

Compounding invariant: existing pages extended, none overwritten
Max 8 outbound links: satisfied on all updated pages
Existing pages updated: 3 (within 5-page cap)
Forward-only linking: no backfill
Tags: reused existing taxonomy
Provenance: frontmatter sources + inline ^[...] citations

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts → wikilint: OK — All checks passed (149 pages, scope: prompts)

[2026-08-03] audit | Prompt Wiki content maintenance closed

Reconciled the historical implementation plan with the current repository state; the plan is complete and remains as historical execution evidence.
Catalog verified: 147 content pages (61 entities, 75 concepts, 5 comparisons, 6 queries).
scripts/wikilint.ps1 -Scope prompts: OK — 147 pages, zero property/link/content errors, zero stale pages, zero broken external links.
Low-confidence findings are intentional provenance labels: three technique summaries sourced from a guide, six leaked or third-party prompt captures, and two vendor/benchmark pages with explicitly bounded claims. They are not unreviewed content gaps.
Historical drift was resolved by updating the plan and catalog counts; no source-to-page implementation gap remains in the current scope.

prompts-log.md — Chronological Action Log

Append-only record of all prompt wiki operations. Format: ## [YYYY-MM-DD] action | subject

[2026-07-16] create | Prompt wiki initialized

Created prompts-log.md — action log for prompt wiki
Created prompts-index.md — content catalog
First ingest: Perplexity AI system prompt (leaked)
Source captured: raw/prompts/leaked/perplexity-ai-system-prompt.md
Entities created: perplexity-system-prompt, perplexity-ai
Concepts created: two-tier-search-architecture

[2026-07-17] ingest | Prompt Engineering Guide (promptingguide.ai/pt)

Raw capture: Fetched 121 pages from promptingguide.ai/pt via 5 parallel task agents
- Introduction (7 pages), Techniques (18), Applications (10), Prompt Hub (25), Models (22), Risks (4), Research (14), Agents (8), Guides (6), Other (7)
Raw sources: Saved to raw/prompts/articles/promptingguide-pt-*.md
Wiki pages created (48):
- Entity: dair-ai — DAIR.AI organization
- Entities (21): FLAN through Model Collection — model-specific pages
- Concepts (11): guide structure, techniques hub, applications, risks, prompt hub, models guide, research, agents, guides, settings
- Concepts (16): zero-shot through graph-prompting — individual technique pages
- Concept: rag — copied with updated path (cross-wiki RAG content)
Updated: prompts-index.md (24 entities, 28 concepts)
Retroactive fix: Pages were initially created in main wiki (wiki/). Moved to wiki/prompts/ after catching the multi-wiki routing error.

[2026-07-22] index | Concept catalog sync (28→44)

Discovered 16 concept files on disk not listed in prompts-index.md
All 16 had full content (none stubs) — lifecycle, ≥2 wikilinks, substantial text
Added to index alphabetically among existing 28
Total concept entries: 28 → 44

Affected pages: Prompt Wiki Catalog.

[2026-07-22] index | Comparison catalog sync (0→5)

Discovered 5 comparison files on disk not listed in prompts-index.md
All 5 had full content (none stubs) — lifecycle, ≥2 wikilinks, substantial text
Added Comparisons section to index with 5 entries
Total page count: 83 → 88

Affected pages: Prompt Wiki Catalog.

[2026-08-03] query | Vault coverage of system prompts (theory + real captures)

Cross-wiki survey: route-classifier ambiguous (0.539); scoped to prompts route since main wiki has no dedicated system-prompt pages.
Summarized: how-to-structure-a-system-prompt (five-section theory), envelope/architecture/technique concepts, safety/defense pages, and 10 leaked real system prompts in raw/prompts/leaked/ with entities claude-code-opus-5-system-prompt, claude-opus-5-system-prompt, grok-personas, grok-expert, grok-4-5-system-prompt, grok-build, perplexity-system-prompt, anthropic-research-instructions, gemini-deep-research-prompt.
scripts/query-check.ps1 -Scope prompts -Catalog → query-check: OK (9/9 checks passed).
No pages created or modified.

[2026-07-22] audit | Remaining raw source audit

124 raw files total: 111 already ingested, 1 duplicate (taxonomy-synthesis.md = taxonomy-synthesis-2026-07-16.md), 12 small peripheral meta-pages (<2KB each: course listings, datasets, tool refs, papers/readings indices)
No substantive new sources found — all high-value content already ingested
Task complete: nothing to ingest

[2026-07-22] lint | Final prompt wiki lint

scripts/wikilint.ps1 -Scope prompts — 91 pages, 9/9 checks PASS
3 pre-existing low-confidence warnings (graph-prompting, prompt-chaining, reflexion) — by design
No broken links, no orphans, no stale pages, all have lifecycle

[2026-07-22] query | Voice/style prompting techniques

Query: “Which prompt engineering techniques can I use to make this prompt better?” — user provided a literary voice/style system prompt
Route: Classifier returned ambiguous → user confirmed prompt wiki
Sources consulted (15 pages): taxonomy-synthesis-2026-07-16.md, promptingguide-pt-introduction-tips.md, promptingguide-pt-introduction-basics.md, promptingguide-pt-introduction-elements.md, promptingguide-pt-introduction.md, promptingguide-pt-guides-optimizing-prompts.md, promptingguide-pt-risks-adversarial.md, promptingguide-pt-prompts-creativity-infinite-primes.md, promptingguide-pt-prompts-creativity-rhymes.md, promptingguide-pt-prompts-adversarial-prompting-jailbreaking-llms.md
Wiki pages read (10): communication-style-spectrum, paired-thinking-response, xml-envelope-format, markdown-header-format, xml-vs-markdown-envelopes, monolithic-prompt-limits, prompt-engineering-techniques, prompt-engineering-guide, prompt-engineering-risks, how-to-structure-a-system-prompt, llm-settings-and-parameters, prompt-engineering-research-gaps, anti-prompt-injection, prompt-hub-catalog
Created: how-to-improve-a-voice-style-prompt (query) — 8 techniques with full citation map
Updated: prompts-index.md — added Queries section (6 entries), correcting oversight
Gap registered: Voice/style prompting is an underserved area in the corpus — zero examples in the ~52 tool taxonomy

[2026-07-31] ingest | Grok Personas leaked system-prompt collection

Source

raw/prompts/leaked/grok-personas.md — leaked collection of 6 xAI/Grok persona system prompts from asgeirtj/system_prompts_leaks (xAI/grok-personas.md)
Captured via webfetch from https://raw.githubusercontent.com/asgeirtj/system_prompts_leaks/refs/heads/main/xAI/grok-personas.md
Evidence body preserved exactly as fetched; provenance frontmatter records source URL and source_type: leaked-prompt

Route validation

scripts/route-classifier.ps1 -Query "grok-personas.md system prompt leak from xAI"
Route: ambiguous (confidence: 0) — no trained match found
Manual override: prompt wiki — this is a real system prompt (leaked), which matches the prompt wiki trigger (“real system prompts, leaked”) from root AGENTS.md Multi-Wiki Routing HARD RULE
Override justified: the classifier has no training data for this specific pattern, but the content is unambiguously a leaked system prompt collection → prompt wiki per wiki/prompts/AGENTS.md

Pages created (1 entity)

wiki/prompts/entities/grok-personas.md — entity documenting the 6 Grok personas (Companion, Unhinged Comedian, Loyal Friend, Homework Helper, Not a Doctor, Not a Therapist); prompt_type: system, envelope_format: markdown, tool_origin: grok; lifecycle: l1, confidence: high

Pages updated (1 entity)

wiki/prompts/entities/grok-1.md — added outbound wikilink to grok-personas to prevent orphan status (forward-only linking); lifecycle promoted l0→l1 (now 2+ sources via persona cross-reference); updated: 2026-07-31

prompts-index.md — added grok-personas entity entry; page count 94→95 (38 entities)
prompts-log.md — this entry

Wikilinks

New page has 6 outbound wikilinks: grok-1, perplexity-system-prompt, claude-code-system-prompt, communication-style-spectrum, anti-prompt-injection, monolithic-prompt-limits

Guardrails applied

Entity threshold (2+ mentions): grok-personas is central to one source (the leaked file itself)
Max 8 outbound links per page: 6 used
Min 2 outbound links per page: 6 used
Raw evidence immutable: body preserved exactly
Forward-only linking: no backfill of sources into grok-1 frontmatter
All tags in prompt wiki taxonomy: prompt-engineering, system-prompt, communication-style, safety-defense, architecture-pattern
Provenance: frontmatter sources + inline ^[…] citations present

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | Grok Expert multi-agent team-leader system prompt

Source

raw/prompts/leaked/grok-expert.md — leaked system prompt for Grok as team leader of a multi-agent collaboration (Harper, Benjamin, Lucas) from asgeirtj/system_prompts_leaks (xAI/grok-expert.md)
Captured via webfetch from https://raw.githubusercontent.com/asgeirtj/system_prompts_leaks/refs/heads/main/xAI/grok-expert.md
Evidence body preserved exactly as fetched; provenance frontmatter records source URL and source_type: leaked-prompt

Route validation

scripts/route-classifier.ps1 -Query "grok-expert.md system prompt leak from xAI multi-agent team leader"
Route: prompts (confidence: 0.691) — trained match, no override needed
Top prompts match: Grok Personas; top main match: Context Engineering

Pages created (1 entity)

wiki/prompts/entities/grok-expert.md — entity documenting Grok Expert: identical-prompt multi-agent team (leader + Harper/Benjamin/Lucas), 13-tool suite, leader-only render components, shared grok-4.5-style safety rules, response-style override slot; prompt_type: system, envelope_format: markdown, tool_origin: grok; lifecycle: l1, confidence: high

Pages updated (2 entities)

wiki/prompts/entities/grok-1.md — added outbound wikilink to grok-expert in the link bar (forward-only linking)
wiki/prompts/entities/grok-personas.md — added grok-expert cross-reference to Related section (sibling leaked Grok prompt)

prompts-index.md — added grok-expert entity entry; page count 95→96 (39 entities)
prompts-log.md — this entry

Wikilinks

New page has 7 outbound wikilinks: grok-1, grok-personas, multi-file-agent-architecture, perplexity-system-prompt, claude-code-system-prompt, communication-style-spectrum, anti-prompt-injection

Guardrails applied

Entity threshold (2+ mentions): grok-expert is central to one source (the leaked file itself)
Max 8 outbound links per page: 7 used
Min 2 outbound links per page: 7 used
Raw evidence immutable: body preserved exactly
Forward-only linking: only added forward links from grok-1 and grok-personas; no backfill
All tags in prompt wiki taxonomy: prompt-engineering, system-prompt, architecture-pattern, safety-defense, communication-style
Provenance: frontmatter sources + inline ^[…] citations present

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] query | Audit and enhance token-reduction catalogue prompt

Query: “Audit this prompt against our knowledge database and enhance it for better results” — user’s prompt asks for a catalogue of token-reduction techniques for vibe coding
Route: Classifier returned ambiguous (main 9.91 / prompts 9.89) → user confirmed prompt wiki
Sources consulted (raw): promptingguide-pt-guides-reasoning-llms.md, promptingguide-pt-applications-context-caching.md, taxonomy-synthesis-2026-07-16.md, promptingguide-pt-models-kimi-k2.5.md
Wiki pages read (prompt wiki): llm-settings-and-parameters, communication-style-spectrum, monolithic-prompt-limits, kimi-k2.5, which-envelope-format-should-i-use, prompt-engineering-research-gaps, xml-vs-markdown-envelopes, template-variable-prompts, cooldown-heuristics, multi-file-agent-architecture, claude-code-system-prompt, paired-thinking-response
Cross-wiki read: wiki/concepts/context-engineering.md (main wiki)
Audit findings delivered: 8 prompt weaknesses + vault-grounded technique inventory for the token-reduction catalogue; enhanced prompt delivered inline (not filed)
Gap registered: No dedicated prompt-wiki concept or query page for token-reduction / context-budget management — token efficiency knowledge is distributed across envelope-format, architecture, and communication-style pages

[2026-07-31] ingest | Grok Build agentic coding CLI system prompt

Source

raw/prompts/leaked/grok-build.md — leaked system prompt for xAI’s Grok Build terminal-native agentic coding assistant (Grok 4.5) from asgeirtj/system_prompts_leaks (xAI/grok-build.md)
Captured via webfetch from https://raw.githubusercontent.com/asgeirtj/system_prompts_leaks/refs/heads/main/xAI/grok-build.md
Evidence body preserved exactly as fetched; provenance frontmatter records source URL and source_type: leaked-prompt

Route validation

scripts/route-classifier.ps1 -Query "Capture and ingest the leaked xAI Grok Build system prompt from asgeirtj/system_prompts_leaks GitHub repository"
Route: ambiguous (confidence: 0.614; main score: 10.5487, top main match Nowledge Mem; prompts score: 16.7535, top prompts match Grok Personas)
Manual override: prompt wiki — user explicitly confirmed prompt wiki after classifier ambiguity. Override justified: leaked real system prompt for an agentic coding tool matches the prompt wiki trigger (“real system prompts, leaked”), and the prompt score (16.75) is well above the main score (10.55) with the top prompt-wiki match being the sibling Grok Personas page.

Pages created (1 entity, 1 concept)

wiki/prompts/entities/grok-build.md — entity documenting Grok Build: 3-part prompt structure (core tag-block behavioral rules, 27-tool JSON schema suite, runtime-injected context templates); prompt_type: system, envelope_format: markdown, tool_origin: grok; lifecycle: l1, confidence: high
wiki/prompts/concepts/action-safety-tiering.md — concept distilling the <action_safety> risk-tiered confirmation policy (reversibility × reach; “one approval is not a blank check”; investigate-unexpected-state tenet); lifecycle: l1, confidence: high

Pages updated (3 entities)

wiki/prompts/entities/grok-1.md — added outbound wikilink to grok-build in the link bar (forward-only linking)
wiki/prompts/entities/grok-personas.md — added grok-build cross-reference to Related section (sibling leaked Grok prompt)
wiki/prompts/entities/grok-expert.md — added grok-build cross-reference to Related section (sibling leaked Grok prompt)

prompts-index.md — added grok-build entity + action-safety-tiering concept entries; page count 96→98 (40 entities, 45 concepts)
prompts-log.md — this entry

Wikilinks

grok-build has 7 outbound wikilinks: grok-1, grok-expert, grok-personas, claude-code-system-prompt, action-safety-tiering, skills-directory-gate, multi-file-agent-architecture
action-safety-tiering has 4 outbound wikilinks: grok-build, plan-before-implement-gate, cooldown-heuristics, anti-prompt-injection

Guardrails applied

Entity threshold (2+ mentions): grok-build is central to one source (the leaked file itself)
Max 8 outbound links per page: 7 (grok-build), 4 (action-safety-tiering)
Min 2 outbound links per page: satisfied on both new pages
Existing pages updated: 3 (within the 5-page batch cap)
Raw evidence immutable: body preserved exactly
Forward-only linking: only added forward links from grok-1, grok-personas, grok-expert; no backfill
All tags in prompt wiki taxonomy: prompt-engineering, system-prompt, architecture-pattern, safety-defense, workflow-pattern, tool-origin
Provenance: frontmatter sources + inline ^[…] citations present
No new tags introduced; no _schema changes needed

Open gap (noted, not created)

raw/prompts/leaked/grok-4.5-system-prompt.md (the base Grok 4.5 safety-rules leak) still has no wiki entity page — out of scope for this ingest, flagged for a future pass.

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | OpenAI Reasoning Models guide

Source

raw/prompts/articles/openai-reasoning-models-guide.md — official OpenAI Reasoning Models guide from developers.openai.com
Captured via webfetch from https://developers.openai.com/api/docs/guides/reasoning
Evidence body preserved exactly as fetched (792 lines); provenance frontmatter records source_url and source_type: web

Route validation

scripts/route-classifier.ps1 -Query "Capture and ingest the official OpenAI Reasoning Models guide from developers.openai.com (reasoning effort, reasoning mode, reasoning tokens, reasoning context, phase parameter)"
Route: ambiguous (confidence: 0.502; main score: 10.8017, top main match RAG vs Fine-tuning; prompts score: 10.9054, top prompts match Prompt Engineering Practical Guides)
Manual override: prompt wiki — user explicitly confirmed prompt wiki after classifier ambiguity. Override justified: the guide is API-level prompt/inference guidance for reasoning models, and the top prompt-wiki match (Prompt Engineering Practical Guides) is the directly on-topic sibling page that this ingest extends.

Pages created (1 entity, 1 concept)

wiki/prompts/entities/gpt-5-6.md — entity documenting the OpenAI GPT-5.6 reasoning model family: model lineup (gpt-5.6 / -sol / -terra / -luna), reasoning effort + mode, reasoning-token budgeting (max_output_tokens, incomplete/max_output_tokens, 25k buffer), reasoning.context, previous_response_id, encrypted_content, summaries, phase; lifecycle: l1, confidence: high
wiki/prompts/concepts/reasoning-effort.md — concept distilling the API-level reasoning control surface (effort scale none→max, mode standard/pro, model-dependent defaults, token budgeting) as inference-level complement to prompt-level CoT; lifecycle: l1, confidence: high

Pages updated (3 pages, within 5-page batch cap)

wiki/prompts/entities/chatgpt.md — added forward link to gpt-5-6 (predecessor → successor)
wiki/prompts/concepts/prompt-engineering-guides.md — Reasoning LLMs section now points to reasoning-effort and gpt-5-6
wiki/prompts/concepts/paired-thinking-response.md — added reasoning-effort cross-reference (user-hidden vs developer-hidden reasoning)

prompts-index.md — added gpt-5-6 entity + reasoning-effort concept entries; page count 98→100 (41 entities, 46 concepts)
prompts-log.md — this entry

Wikilinks

gpt-5-6 has 6 outbound wikilinks: reasoning-effort, chatgpt, gpt-4, prompt-engineering-guides, llm-models-guide, paired-thinking-response
reasoning-effort has 3 outbound wikilinks: chain-of-thought, paired-thinking-response, gpt-5-6

Guardrails applied

Entity threshold (2+ mentions): gpt-5.6 family is central to the source
Max 8 outbound links per page: 6 (gpt-5-6), 3 (reasoning-effort)
Min 2 outbound links per page: satisfied on both new pages
Existing pages updated: 3 (within the 5-page batch cap)
Raw evidence immutable: body preserved exactly
Forward-only linking: only added forward links from chatgpt, prompt-engineering-guides, paired-thinking-response; no backfill
Tags reused from existing taxonomy (modelo, openai, reasoning, llm; prompt-engineering, instruction-pattern); no new tags introduced; no _schema changes needed
Provenance: frontmatter sources + inline ^[…] citations present
No contradictions with existing pages (reasoning-effort positioned as complement to chain-of-thought, not replacement)

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | OpenAI Structured Outputs Guide + OpenAI Prompt Engineering Guide

Raw capture

raw/prompts/articles/openai-structured-outputs-guide.md — official OpenAI Structured Outputs guide (developers.openai.com), fetched via webfetch; evidence body immutable
raw/prompts/articles/openai-prompt-engineering-guide.md — official OpenAI Prompt Engineering guide (developers.openai.com), fetched via webfetch; evidence body immutable

Route validation

scripts/route-classifier.ps1 -Query "Capture and ingest the official OpenAI Structured Outputs guide from developers.openai.com (structured outputs, response_format, JSON schema)" → Route: ambiguous (confidence 0.521; main score 9.0425, top main match Embeddings; prompts score 9.8227, top prompts match Prompt Engineering Applications)
scripts/route-classifier.ps1 -Query "Capture and ingest the official OpenAI Prompt Engineering guide from developers.openai.com (prompt engineering techniques and best practices)" → Route: ambiguous (confidence 0.527; main score 12.2207, top main match RAG vs Fine-tuning; prompts score 13.6429, top prompts match Prompt Engineering Applications)
Manual override: prompt wiki — user explicitly confirmed both sources route to the prompt wiki after classifier ambiguity. Override justified: both are official API-level prompt/inference guides on-topic for the prompt wiki, and their closest existing siblings (prompt-engineering-guides, gpt-5-6) live there.

Pages created (2 concepts)

wiki/prompts/concepts/structured-outputs.md — JSON-Schema-guaranteed output via text.format: json_schema; vs JSON mode; explicit refusals; strict schema subset + limits; SDK helpers (Pydantic/Zod); lifecycle l1, confidence high
wiki/prompts/concepts/message-roles.md — developer/user/assistant priority chain; instructions parameter (current-generation scope); developer-message structure (Identity/Instructions/Examples/Context); lifecycle l1, confidence high

Pages updated (3 content pages, within 5-page batch cap)

wiki/prompts/entities/gpt-5-6.md — added Structured Outputs Support section (structured-outputs) and Model Selection section (reasoning vs GPT models, GPT-5 best practices); sources += both new raw paths; 7 outbound links
wiki/prompts/concepts/prompt-engineering-guides.md — added OpenAI Prompt Engineering Guide and OpenAI Structured Outputs Guide sections; sources += both raw paths; lifecycle l0→l1; 8 outbound links
wiki/prompts/concepts/llm-settings-and-parameters.md — added Structured Outputs and Determinism section (schema guarantees vs sampling parameters); sources += structured-outputs raw path; lifecycle l0→l1

prompts-index.md — added message-roles + structured-outputs; page count 100→102 (41 entities, 48 concepts)
prompts-log.md — this entry

Guardrails applied

Compounding invariant: all three existing pages preserved and extended, none overwritten
Entity/concept threshold: both new concepts are central to their source
Max 8 outbound links: gpt-5-6 7, prompt-engineering-guides 8, llm-settings-and-parameters 4, structured-outputs 4, message-roles 5
Min 2 outbound links: satisfied on all five pages
Existing pages updated: 3 (within 5-page cap)
Forward-only linking: no backfill
Tags: reused existing vocabulary (prompt-engineering, instruction-pattern, openai, llm); no new tags; no _schema changes
Provenance: frontmatter sources + inline ^[…] citations
Contradictions: none — message-roles and structured-outputs positioned as complements to existing pages (chain-of-thought, reasoning-effort)

Open questions carried to pages

Refusal behavior under reasoning effort; schema adherence across multi-step agentic loops (structured-outputs)
developer-role interaction with tool-result messages; Chat Completions parity (message-roles)
Dated claim: OpenAI reusable prompt objects (v1/prompts) shutdown November 30, 2026

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] capture | Claude Code costs — manage costs and reduce token usage

Raw capture only: raw/prompts/articles/claude-code-costs-reduce-token-usage.md — official Claude Code docs page (code.claude.com/docs/en/costs), fetched via webfetch; evidence body immutable
Route validation: scripts/route-classifier.ps1 -Query "Capture and ingest the Claude Code docs page on costs and reducing token usage (token consumption reduction, prompt caching, context management, agent configuration)" → Route: ambiguous (confidence 0.515; main score 16.8941, top main match Claude Code; prompts score 15.9102, top prompts match Prompt Engineering Research Gaps)
Manual override: prompt wiki — user explicitly selected prompt wiki. Override justified: token-usage reduction for a coding agent is prompt/context-engineering product guidance (tool-origin: claude-code).
Status: captured; step-3 analysis and page synthesis pending user confirmation

[2026-07-31] ingest | Claude Code costs — manage costs and reduce token usage

Raw capture

raw/prompts/articles/claude-code-costs-reduce-token-usage.md — official Claude Code docs page (code.claude.com/docs/en/costs); evidence body immutable

Route validation

scripts/route-classifier.ps1 -Query "Capture and ingest the Claude Code docs page on costs and reducing token usage (token consumption reduction, prompt caching, context management, agent configuration)" → Route: ambiguous (confidence 0.515; main score 16.8941, top main match Claude Code; prompts score 15.9102, top prompts match Prompt Engineering Research Gaps)
Manual override: prompt wiki — user explicitly selected prompt wiki. Override justified: token-usage reduction for a coding agent is prompt/context-engineering product guidance (tool-origin: claude-code).

Pages created (1 concept)

wiki/prompts/concepts/token-usage-reduction.md — context management, prompt caching, model selection, extended-thinking and agent-team cost levers; lifecycle l1, confidence high

Pages updated (4 content pages, within 5-page batch cap)

wiki/prompts/entities/claude-code-system-prompt.md — added Costs and Token Usage section (/usage, /usage-credits, caching, compaction, model selection); sources += claude-code costs raw; lifecycle l0→l1
wiki/prompts/entities/claude-3.md — added Prompt Caching e Seleção de Modelo section (Sonnet default / Opus / haiku, cache lifetime 1h vs 5min); sources += claude-code costs raw; lifecycle l0→l1
wiki/prompts/concepts/llm-settings-and-parameters.md — added Extended Thinking Settings section (/effort, MAX_THINKING_TOKENS, adaptive vs fixed budgets); sources += claude-code costs raw
wiki/prompts/concepts/ai-agents-guide.md — added Agent Team Token Costs section (~7× tokens in plan mode, team-size scaling); sources += claude-code costs raw

prompts-index.md — added token-usage-reduction; page count 102→103 (41 entities, 49 concepts)
prompts-log.md — this entry

Guardrails applied

Compounding invariant: all four existing pages preserved and extended, none overwritten
Max 8 outbound links: token-usage-reduction 5, claude-code-system-prompt 6, claude-3 7, llm-settings-and-parameters 6, ai-agents-guide 5
Min 2 outbound links: satisfied on all five pages
Existing pages updated: 4 (within 5-page cap)
Forward-only linking: no backfill
Tags: reused existing vocabulary (prompt-engineering, claude-code, context-engineering, anthropic); no new tags; no _schema changes
Provenance: frontmatter sources + inline ^[…] citations
Contradictions: none — token-usage-reduction positioned as complement to llm-settings-and-parameters and ai-agents-guide

Open questions carried to page

Cache-lifetime distinction (1h vs 5min) is version-sensitive
model: haiku quality on non-trivial subagent tasks

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] query | Prompt techniques for a repo tech-debt housekeeping prompt (prompt-master)

Query: “Query the wiki to find prompt techniques to use alongside prompt-master before delivering a tech-debt housekeeping prompt for this repo”
Route: scripts/route-classifier.ps1 returned ambiguous (confidence 0.554; main 11.61 / prompts 9.35) → user confirmed prompt wiki
Technique inventory delivered (7 pages read): action-safety-tiering (reversibility × reach confirmation policy), plan-before-implement-gate (analyze→plan→approve→implement), preliminary-context-gathering (inventory before acting), task-state-machine (pending→active→done→verified artifact lifecycle), communication-style-spectrum (explain-before-do + enumerate-exhaustively for the classification report), cooldown-heuristics (loop guard), chain-of-thought-pattern (step-by-step reasoning)
Cross-wiki grounding: vault guardrails (raw/ immutability, compounding invariant, log every action) applied to the generated prompt
Output: Claude Code housekeeping prompt delivered inline (not filed); target tool and repo scope confirmed with user

[2026-07-31] ingest | Batch 1 — token-compression tools (RTK, Headroom, Token Optimizer MCP)

Task: Test batch ingest mode — capture and ingest every link inside D:\tokenreduction\techniques-with-links.md
Route: scripts/route-classifier.ps1 returned ambiguous (confidence 0.502; main 6.023 / prompts 6.059) → user confirmed prompt wiki
Source doc provenance: raw/prompts/token-reduction-techniques.md (captured copy of the doc, immutable)
Batch capture: 53 unique URLs extracted from the doc → external_links.py capture --batch → 48 queued (3 HTTP 403, no snapshot: help.openai.com/en/articles/11096431, openai.com/index/unrolling-the-codex-agent-loop, platform.openai.com/docs/api-reference/chat/create; 2 redirects collapsed as duplicates)
Batch 1 claimed (5, token-compression tools theme):
- RTK repo + site + issue #582 (github-com-rtk-a2ac6abb.md, rtk-ai-app-index-1d38dd66.md, github-com-f7d4988e.md)
- Headroom (github-com-headroom-cbd05f4e.md)
- Token Optimizer MCP (github-com-token-optimizer-mcp-987700a8.md)
Candidates: 3 written to raw/.ingest-queue/candidates/, all approved by user (--decide approve)
Entities created (3): rtk-rust-token-killer (contested: true), headroom, token-optimizer-mcp
Existing pages updated (1): token-usage-reduction — recorded RTK contradiction (input compression can increase cost: +18% cost / 50% more output tokens, issue #582); contested: true
Key finding (contradiction pass): token reduction ≠ cost reduction — output tokens billed ~5× input on Opus-class; input-only proxies can push cost negative (RTK #582). Headroom adds output shaping; Token Optimizer MCP measures against control arm.
Sources completed: all 5 (--complete) after materialization
Backlog remaining: 43 pending (next batches to claim)
Compounding invariant: token-usage-reduction preserved and extended, not overwritten
Tags: token-usage-reduction used as tag consistent with sibling concept page; no _schema.md changes
Contradictions: rtk-rust-token-killer ↔ token-usage-reduction (both contested: true, linked in contradictions frontmatter)

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | Batch 4 — Claude Code official docs (5 sources)

Task: continue batch ingest of all links from D:\tokenreduction\techniques-with-links.md (backlog processing)
Batch 4 claimed (5, Claude Code official docs):
- docs-anthropic-com-best-practices-855abde4.md (Best practices for Claude Code)
- code-claude-com-model-config-815ffab5.md (Model configuration)
- code-claude-com-agents-47001d98.md (Run agents in parallel)
- code-claude-com-context-window-4fb02d1e.md (Explore the context window)
- code-claude-com-sub-agents-5d62f063.md (Create custom subagents)
Candidates: 5 written, all approved (5 create, 0 merge)
Concepts created (5): claude-code-best-practices (verification levels, explore→plan→code→commit, context-window primacy), claude-code-model-configuration (provider-dependent aliases, opusplan hybrid, availableModels governance), claude-code-parallel-agents (four parallelization approaches, worktree isolation, token multiplier), claude-code-context-window (startup context, compaction survival, 1M-token gate, /context), claude-code-subagents (markdown+frontmatter, five-scope precedence, Haiku routing, background tool pool)
Existing pages updated: 0 (all new pages)
Sources completed: all 5 (--complete)
Backlog remaining: 28 pending
Compounding invariant: all new pages; cross-links between claude-code-subagents and claude-code-parallel-agents both new this batch (forward-only preserved)
Contradictions: none new; version-sensitive claims (v2.1.153–v2.1.219 gates, Opus 5/Sonnet 5/Fable 5 model names) annotated in Open Questions; provider-dependent alias resolution flagged as seeming-contradiction source
Tags: prompt-engineering, workflow-pattern, configuration, architecture-pattern, tool-origin reused from existing vocabulary

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | Batch 2 — Claude Code cost/context official docs (5 sources)

Task: continue batch ingest of all links from D:\tokenreduction\techniques-with-links.md (backlog processing)
Batch 2 claimed (5, Claude Code official docs):
- code-claude-com-costs-ea267c0e.md (Manage costs effectively)
- platform-claude-com-prompt-caching-a167718a.md (Prompt caching)
- code-claude-com-memory-09a26217.md (How Claude remembers your project)
- code-claude-com-sessions-0ab9c7fd.md (Manage sessions)
- code-claude-com-env-vars-12397363.md (Environment variables)
Candidates: 5 written, all approved (4 create + 1 merge into token-usage-reduction)
Concepts created (4): prompt-caching (0.1× reads / 1.25–2× writes, model-specific prefix minimums), claude-code-memory (CLAUDE.md hierarchy + auto memory), claude-code-sessions (resume/branch/export), claude-code-environment-variables (token budgets, caching, model pins)
Existing pages updated (1): token-usage-reduction — added rate-limit tiers, /usage attribution, background usage, CLAUDE.md→skills migration; promoted l1→l2
Sources completed: all 5 (--complete)
Backlog remaining: 38 pending
Compounding invariant: token-usage-reduction preserved and extended
Contradictions: none new (prompt-caching numbers annotated as version-sensitive in Open Questions)
Tags: claude-code, memory, workflow-pattern, configuration reused from existing vocabulary

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | Batch 3 — context-reduction research papers + codebase-memory repo (5 sources)

Task: continue batch ingest of all links from D:\tokenreduction\techniques-with-links.md (backlog processing)
Batch 3 claimed (5, research cluster):
- arxiv-org-2603-29919-0aae2e40.md (SkillReducer, arXiv 2603.29919)
- arxiv-org-2607-12161-a5d415f4.md (Token Reduction Is Not Cost Reduction, arXiv 2607.12161)
- arxiv-org-2607-01916-96bad417.md (ContextSniper, arXiv 2607.01916)
- arxiv-org-2603-27277-af87911a.md (Codebase-Memory paper, arXiv 2603.27277)
- github-com-codebase-memory-mcp-f33fbe15.md (codebase-memory-mcp repo by DeusData)
Candidates: 4 written (paper+repo merged into one page), all approved (4 create, 0 merge)
Concepts created (4): skill-reducer (48% desc / 39% body compression, +2.8% quality), token-reduction-is-not-cost-reduction (−38% tool-output → +6.8% billed cost; cache ≈80–87% of bill), contextsniper (−51.5%/−36.4% and −38.9%/−27.3%, resolution unchanged, recoverable evidence), codebase-memory (Tree-Sitter KG over MCP; 83% vs 92% quality at ~10× fewer tokens; 158 languages; aliases: codebase-memory-mcp)
Existing pages updated: 0 (all new pages)
Sources completed: all 5 (--complete)
Backlog remaining: 33 pending
Compounding invariant: all new pages; rtk-rust-token-killer and token-usage-reduction cross-linked but not modified (forward-only linking preserved)
Contradictions: recorded the paper/repo token-claim discrepancy (~10× vs 99.2%, different baselines) in codebase-memory Open Questions; TRINCR paper independently corroborates RTK issue #582 (prompt-cache ~87% of cost)
Tags: prompt-engineering, token-usage-reduction, research-gap, tool-origin reused from existing vocabulary

Verification

pwsh -File scripts/wikilint.ps1 -Scope prompts — see output below

[2026-07-31] ingest | Batch 5 — external-link captures: cache/context/cost tools (5 sources)

Task: continue batch ingest of the external-link capture backlog (raw/external/) from the batch workflow
Batch 5 claimed (5, token/context-cost cluster):
- github-com-claude-code-cache-fix-0e94ee20.md (claude-code-cache-fix)
- github-com-codeburn-99fcb193.md (codeburn)
- github-com-contextsniper-7749d2f0.md (ContextSniper GitHub repo — MERGE into existing contextsniper)
- github-com-context-mode-052ecc22.md (Context Mode)
- github-com-crush-fb00f031.md (Crush)
Candidates: 5 written (4 create + 1 merge), all approved (4 create, 1 merge)
Entities created (4): claude-code-cache-fix (silent /resume cache break fixed by local proxy; “up to 20x cost”; cache-hit observability), codeburn (local-first usage tracker, 36 tools, never intercepts; optimize/guard/compare), context-mode (sandboxed tool output ~98% claimed; FTS5 session memory; hook-enforced routing), crush (charmbracelet multi-model CLI; crushrc/AGENTS.md; FSL-1.1-MIT)
Existing page merged (1): contextsniper — extended with repo Implementation section (dual plugins, repo numbers, config) + 2 new Open Questions (authorship unverified, number discrepancy)
Sources completed: all 5 (--complete)
Backlog remaining: 28 pending
Compounding invariant: contextsniper preserved and extended, not overwritten
Contradictions: recorded codeburn 31-vs-36 internal count; Context Mode 17-vs-18 platforms; ContextSniper repo-vs-paper token/cost number discrepancies
Tags: prompt-engineering, tool-origin, token-usage-reduction reused from existing vocabulary
Orphan resolution (forward-only, guardrail #6): 4 new entity pages had zero inbound links. Per vault precedent (log entry 2026-07-31 grok-1→grok-personas), existing pages were updated to point forward at the new pages — no backfilling, 4 pages updated (total 5 with contextsniper, exactly at the max-5 budget):
- prompt-caching → +claude-code-cache-fix (cache-hit repair)
- claude-code-sessions → +codeburn (reads on-disk sessions)
- contextsniper → +context-mode (select-don’t-destroy parallel)
- ai-agents-guide → +crush (agentic CLI instance)
Verification: pwsh -File scripts/wikilint.ps1 -Scope prompts → OK — all checks passed (121 pages, 0 errors). Queue: 0 claimed, 0 failed, 23 pending.

[2026-07-31] ingest | Batch 6 — Agent Skills ecosystem (5 sources)

Task: continue batch ingest of the external-link capture backlog (raw/external/) from the batch workflow
Batch 6 claimed (5, Agent Skills ecosystem cluster):
- ponytailskill-com-718d054e.md (ponytail landing page)
- github-com-ponytail-d8b2056c.md (ponytail GitHub repo — MERGED with landing into one entity)
- skills-re-cc5c6e1d.md (skills.re registry)
- openagentskill-com-6cae86ca.md (OpenAgentSkill registry + recommendation API)
- github-com-anthropic-best-practices-md-d3c57d61.md (Anthropic SKILL.md authoring best practices, mirror in davila7/claude-code-templates)
Route validation: batch query → ambiguous (confidence 0.532; main 11.6677, prompts 13.2449); each subsource re-validated individually (best-practices: main 9.5133/prompts 15.7628 conf 0.624; ponytail: main 11.398/prompts 12.6153 conf 0.525; skills.re: main 16.7452/prompts 16.0239 conf 0.511; OpenAgentSkill: main 9.5099/prompts 10.5119 conf 0.525) → user confirmed all 5 to prompt wiki after ambiguity
Candidates: 4 written (3 create + 1 landing/repo merge into ponytail), all approved (4 create, 0 merge)
Pages created (1 concept, 3 entities):
- skill-authoring-best-practices (concept) — concise SKILL.md, third-person descriptions, degrees of freedom high/medium/low, progressive disclosure one-level-deep; ties to skills-directory-gate, skill-reducer
- ponytail (entity, aliases: ponytail-skill) — anti-overbuild skill; YAGNI ladder; agentic benchmark −54% LOC / −22% tokens / −20% cost / 100% safe; portable to 20+ agents
- skills-re (entity, aliases: the-registry-of-agent-skills) — open registry; “a skill is an instruction, not an install”; CLI install per agent; public diffable reviews; ~7,380 skills in 8 categories
- openagentskill (entity, aliases: oas) — registry + agent-facing API (POST /api/agent/resolve → skill + alternatives + fit/risk + install plan); 20,842 skills, 104 agent surfaces; compares to skills.sh / agentskills.io / native docs
Existing pages updated: 0 (all new pages)
Sources completed: all 5 (--complete)
Backlog remaining: 18 pending (queue prior to batch)
Compounding invariant: all new pages; registry pair (skills-re, openagentskill) and skill-authoring-best-practices cross-link each other; ponytail links forward to skill-authoring-best-practices and skills-directory-gate without modifying them
Contradictions (recorded in pages): ponytail benchmark corrects the earlier 80–94% LOC-reduction claim (issue #126: bare-model baseline padded answers; agentic mean −54% is the defensible figure); terse-reasoning models may reverse (thinking tokens); caveman control arm referenced in plain text (no page). Cross-source: OpenAgentSkill 20,842 vs skills.re ~7,380 indexed skills — different inclusion criteria, flagged not comparable
Tags: prompt-engineering, workflow-pattern, architecture-pattern, tool-origin reused from existing vocabulary; no _schema.md changes
Verification: pwsh -File scripts/wikilint.ps1 -Scope prompts → see output below

[2026-07-31] lint | synaptic-lint — scope all

Structural command: pwsh -File scripts/wikilint.ps1 -Scope all → OK, 219 pages; no Property or Link errors.
Prompt semantic findings: 14 raw/prompts/ sources currently have no contributing page, and 25 prompt pages cite raw/external/ rather than the prompt semantic root raw/prompts/.
Four prompt pages have low-confidence findings in the structural advisory layer; no automatic repairs were made.
[2026-07-31] semantic-repair | Prioridades 1–3 — Snapshots capturados em raw/external/ foram reconhecidos como evidência válida para a Prompt Wiki quando preservam metadados de captura; fontes Prompt sem página foram registradas com destinos sugeridos em docs/semantic-data-gaps.md. Nenhuma página foi criada por heurística.
[2026-07-31] ingest | Prompt resource backlog — Dez listas do Prompt Engineering Guide foram reunidas em prompt-engineering-resources; o inventário de redução de tokens foi ligado a token-usage-reduction; a síntese taxonômica duplicada foi ligada a prompt-engineering-techniques; e o prompt vazado do Grok 4.5 recebeu uma entidade cautelosa, grok-4-5-system-prompt, com confidence: low e contested: true. Fontes brutas não foram modificadas.

[2026-07-31] ingest | Batch 7 — OpenAI API docs (models, caching, reasoning, structured outputs, prompting)

Task: batch ingest of OpenAI API documentation captures from raw/external/, following the batch workflow.
Batch 7 claimed (5, OpenAI API docs cluster):
- developers-openai-com-models-bd8184ce.md (Models overview: GPT-5.6 family)
- developers-openai-com-prompt-caching-a7f81296.md (Prompt caching: breakpoints, cache keys)
- developers-openai-com-reasoning-bc1c15f2.md (Reasoning models: effort, thinking)
- developers-openai-com-structured-outputs-bd36d078.md (Structured outputs: schema, function calling)
- developers-openai-com-prompting-2cb0df19.md (Prompting: refine, ship, /v1/prompts deprecation)
Route validation: all 5 OpenAI API docs → prompt wiki (confidence high; vendor API documentation for prompt/technique domain).
Candidates: 5 written, all merged into existing pages (0 new pages):
- cand-b7-models → gpt-5-6 (pricing table Sol/Terra/Luna, 128K max output, 1.05M context, Feb 16 2026 cutoff)
- cand-b7-openai-prompt-caching → prompt-caching (GPT-5.6 implicit breakpoint change, prompt_cache_breakpoint/prompt_cache_key, mode: explicit, TTL 30m, write 1.25×, min 1024 tokens)
- cand-b7-openai-reasoning → reasoning-effort (effort best-for table, pro-mode billing, interleaved thinking, prompting advice)
- cand-b7-openai-structured-outputs → structured-outputs (function-calling vs response_format decision rule, supported models, compatibility matrix)
- cand-b7-openai-prompting → prompt-engineering-guides (OpenAI refine/ship rules, v1/prompts deprecation timeline June 3 / Nov 30 2026)
Existing pages updated (5): gpt-5-6, prompt-caching, reasoning-effort, structured-outputs, prompt-engineering-guides — all extended (sources added to frontmatter, inline ^[...] citations, Open Questions).
Compounding invariant: all updates preserved existing content and added new material; no page overwritten.
Contradictions: none recorded.
Sources completed: all 5 (--complete). Queue: 35 done, 13 pending, 0 claimed.
Tags: existing vocabulary reused; no _schema.md changes.
Verification: pwsh -File scripts/wikilint.ps1 -Scope prompts → OK, 127 pages; see output below.

[2026-07-31] lint | Batch 7 pre-flight + orphan/type repair (grok-4-5-system-prompt, prompt-engineering-resources)

Lint before repair reported 5 failures from pages created in the earlier semantic-repair/ingest pass (not Batch 7): orphan grok-4-5-system-prompt, orphan prompt-engineering-resources; unrecognized relationship types instance-of and illustrates (×2) in grok-4-5-system-prompt.
Repairs (forward-only, no page rewrites):
- grok-4-5-system-prompt: "instance-of" → "belongs-to", "illustrates" ×2 → "implements" (vocabulary from _schema.md).
- grok-1: added forward link [[grok-4-5-system-prompt]] (orphan inbound fix).
- prompt-engineering-resources: orphan resolved — inbound link already present from prompt-engineering-guide; stale lint state cleared.
Verification: pwsh -File scripts/wikilint.ps1 -Scope prompts → OK, 127 pages, 0 errors.
[2026-07-31] lint | Fechamento das prioridades 1–3 — A auditoria estrutural em escopo all terminou com wikilint: OK em 225 páginas. O backlog Prompt e os snapshots externos estão roteados; a lacuna histórica da wiki principal foi resolvida por revalidação das páginas, sem fabricar evidência.

[2026-07-31] ingest | Batch 8 — OpenCode cluster (CLI, Config, Compaction, archived GitHub repo)

Task: batch ingest of the OpenCode source cluster from raw/external/, following the batch workflow.
Batch 8 claimed (4, OpenCode cluster):
- dev-opencode-ai-cli-b01a850a.md (OpenCode V2 CLI reference)
- opencode-ai-config-85d6db47.md (OpenCode V2 config reference)
- opencode-ai-compaction-d433d7dd.md (OpenCode V2 compaction)
- github-com-opencode-104f4b54.md (archived opencode-ai/opencode Go repo → moved to Crush)
Route validation: pwsh -File scripts/route-classifier.ps1 -Query "OpenCode CLI, config, compaction and GitHub repository documentation..." → ambiguous (confidence 0.545; main 15.1151, prompts 18.1411, top_main docmd.io, top_prompts “Grok Build (Agentic Coding CLI)”). User overrode to prompt wiki (tool-origin precedent: grok-build, cursor-system-prompt, claude-code-model-configuration).
Critical batch finding (name collision): two different products named “OpenCode” — archived Go opencode-ai/opencode (2025-09-18, continued as crush/charmbracelet) vs current V2 anomalyco/opencode. Documented as a warning on the opencode entity page.
Candidates: 2 written, both approved (2 create, 0 merge):
- cand-b8-opencode → create opencode (entity, tool-origin) — V2 CLI surface, whitelist permission model, V2 config (JSONC, $schema, load order, accepted-but-inert fields), compaction summary; naming-collision warning; links claude/cursor/cline/grok-build/context-compaction/openagentskill/skill-authoring-best-practices
- cand-b8-context-compaction → create context-compaction (concept) — checkpoint = structured summary + retained tail (keep.tokens 8000, buffer 20000); lossy but durable; preflight estimate trigger + one-shot overflow recovery; instruction-epoch decoupling; distinct from claude-code-context-window
Contradictions found across the batch (mandatory pass): ① “OpenCode” = two unrelated products (archived→Crush vs V2); ② auto-compact V1 (95% context, new session) vs V2 (in-place checkpoint, keep.tokens) — both dated positions preserved; ③ config V1 (mcpServers, ~/.opencode.json) vs V2 (mcp.servers, ~/.config/opencode/). Recorded inline; no contradictions cross-links since both positions live on the same page.
Pages created (1 entity, 1 concept): opencode, context-compaction
Existing pages updated: 0
Compounding invariant: new pages only; no existing page overwritten.
Sources completed: pending until lint gate passes.
Tags: prompt-engineering, tool-origin, architecture-pattern, workflow-pattern, research-gap — all existing vocabulary; no _schema.md changes.

[2026-07-31] ingest | Batch 9 — Claude Platform tooling cluster (manage tool context, tool search, batch processing)

Task: batch ingest of the Claude Platform tooling docs from raw/external/, following the batch workflow.
Batch 9 claimed (3, Claude Platform cluster):
- platform-claude-com-manage-tool-context-625d4da8.md (Manage tool context)
- platform-claude-com-tool-search-tool-081b885d.md (Tool search tool)
- platform-claude-com-batch-processing-d8088a83.md (Batch processing)
Route validation: pwsh -File scripts/route-classifier.ps1 -Query "Graphify knowledge-graph skill for Claude Code; Claude Platform docs on tool context, batch processing and tool search; Caveman token-efficient agent-native stack; Vibe-Coder-MCP tool server; OpenAI developers homepage" → ambiguous (confidence 0.523; main 36.089, prompts 32.8822, top_main Graphify, top_prompts Token Optimizer MCP). User overrode to prompt wiki for the entire 9-source backlog.
Candidates: 3 written, all approved (3 create, 0 merge):
- cand-b9-tool-search → create tool-search-tool (concept) — on-demand tool discovery, defer_loading, tool_reference expansion, regex/BM25 variants, MCP default_config, 10k deferred tools limit, ~85% context reduction on multiserver setups
- cand-b9-tool-context → create tool-context-management (concept) — organizing frame for the four approaches (tool search, programmatic tool calling, prompt caching, context editing), adoption ordering, composition
- cand-b9-batch-processing → create message-batch-processing (concept) — Message Batches API, 50% pricing, 100k/256MB limits, 24h expiry, 29-day results, server tools + pause_turn, 300k extended output (beta), cache stacking
Contradictions found across the batch (mandatory pass): none substantiva. Tool search GA naming vs model-preview names in the batch doc reflect naming in flux, not conflict; the three sources cross-reference consistently (tool search ↔ batch requests; batch doc lists tool search among server tools). The sources cover complementary layers.
Pages created (3 concepts): tool-search-tool, tool-context-management, message-batch-processing
Existing pages updated: 0 (Guardrail #5: 0/5 used)
Compounding invariant: new pages only; no existing page overwritten.
Sources completed: pending until lint gate passes.
Tags: prompt-engineering, architecture-pattern, workflow-pattern, token-usage-reduction — all existing vocabulary; no _schema.md changes.

[2026-07-31] ingest | Batch 10 — Graphify cluster (GitHub README v8 + docs quickstart)

Task: batch ingest of the Graphify skill/repo documentation from raw/external/, following the batch workflow.
Batch 10 claimed (2, Graphify cluster):
- github-com-graphify-dc3bc982.md (Graphify-Labs/graphify README, v8; → redirected from safishamsi/graphify)
- graphify-com-docs-e68d0288.md (Graphify docs Quickstart)
Route validation: inherited from Batch 9 backlog route decision (ambiguous 0.523 → user overrode “Prompt wiki p/ tudo”). Graphify is a /graphify skill (tool-origin) precedent like token-optimizer-mcp; the existing main-wiki wiki/entities/graphify.md (created 2026-07-14) is left untouched per the routing decision.
Candidates: 1 written, approved (1 create):
- cand-b10-graphify → create graphify (concept) — /graphify skill, local tree-sitter AST extraction (0 LLM credits), god nodes + Leiden communities + confidence tags (EXTRACTED/INFERRED/AMBIGUOUS), MCP server (10 tools, stdio + shared HTTP), always-on hooks/AGENTS.md install across 20+ assistants, graphifyy PyPI package, benchmarks (LOCOMO recall@10 0.497, 76% LongMemEval), prs --triage/--conflicts.
Contradictions found across the batch (mandatory pass): none. Minor fact divergence (README “20+ assistants” vs docs “17 assistants supported”) attributed to snapshots at different points in time — not a factual conflict; no contested/contradictions flags.
Pages created (1 concept): graphify
Existing pages updated: 0 (Guardrail #5: 0/5 used)
Compounding invariant: new concept page only; main-wiki wiki/entities/graphify.md untouched.
Sources completed: pending until lint gate passes.
Tags: prompt-engineering, tool-origin, architecture-pattern — all existing vocabulary; no _schema.md changes.

[2026-07-31] ingest | Batch 11 — Caveman cluster (caveman.so + INSTALL.md)

Task: batch ingest of the Caveman token-efficient agent stack documentation.
Batch 11 claimed (2, Caveman cluster):
- caveman-so-e91601d5.md (caveman.so homepage: skill + engine + enterprise + research)
- github-com-install-md-5709b2d9.md (INSTALL.md: one-liner + 30-agent install matrix + flags/privacy/troubleshooting)
Route validation: inherited ambiguous backlog decision → prompt wiki (tool-origin precedent token-optimizer-mcp).
Candidates: 1 written, approved (1 create):
- cand-b11-caveman → create caveman (entity, tool_origin: claude-code, output token reduction) — /caveman skill (~65% output cut), caveman wrap byte-safe local proxy, 9-compressor Enterprise engine + eval-gated routing, CaveGemma fine-tune (gemma-4-31B-it, QLoRA r16), install across 30+ agents (curl|bash/irm|iex, --with-init→.opencode/AGENTS.md+.cursor/rules/).
Contradictions found across the batch (mandatory pass): none. Homepage claims (product/engineer) and INSTALL.md (surface/flags/privacy) cover complementary layers; consistent on /caveman, caveman.wrap, npx skills add, .opencode/AGENTS.md.
Pages created (1 entity): caveman
Existing pages updated: 1 — added [[caveman]] forward-link to token-optimizer-mcp Related (peer token-reduction tool-origin, resolves orphan without backfilling stable pages beyond same-ingest new content).
Pages created 0 (Batch 11 new content: 1 entity)
Sources completed: pending until lint gate passes.
Tags: prompt-engineering, tool-origin, token-usage-reduction, workflow-pattern — all existing vocabulary; no _schema.md changes.

Gains / Losses

Gained: Caveman (caveman/openagentskill skill-install pattern + output-token-reduction cluster) documented; orphan for caverman surfaced and fixed via peer forward-link (same-ingest new-content cluster token-optimizer-mcp).
Losses/complexity: +1 forward-link edit on token-optimizer-mcp; token-reduction cluster now has 3 mutually-cross-referenced tool-origin entities (caveman, token-optimizer-mcp, rtk-rust-token-killer) — modest graph density increase.

[2026-07-31] ingest | Batch 12 — vibe-coder-mcp cluster (npm MCP server + OpenAI homepage)

Task: batch ingest of the final claimed pair; finish backlog routing for this ingest cycle.
Batch 12 claimed (2):
- github-com-vibe-coder-mcp-afed2aee.md (c9c40b79772056c9 — vibe-coder-mcp npm package README: 10 MCP tools, RDD task manager, transports, CLI/REPL)
- developers-openai-com-9dcdba4f.md (cc1760193b07a0c7 — developers.openai.com homepage)
Route validation: inherited ambiguous backlog decision → prompt wiki (tool-origin precedent token-optimizer-mcp, caveman).
Candidates: 1 written, approved (1 create):
- cand-b12-vibe-coder-mcp → create vibe-coder-mcp (entity, tool_origin: cursor, prompt_type: product-config) — MCP server extending Cursor/Cline/Claude Desktop; 10 tools (vibe-task-manager RDD 21 intents, research-manager via Perplexity Sonar, map-codebase 35+ languages, generate-prd/user-stories/task-list, run-workflow); 4 transports (stdio/SSE/WS/HTTP); OPENROUTER_API_KEY required; CLI vibe + REPL. README claims 99.9% test success / 95-97% token reduction — marked contested: true, no external benchmark captured.
Discarded (no candidate): developers-openai-com-9dcdba4f.md — pure navigation landing page (API dashboard/ChatGPT/Codex/Docs links + llms.txt index); zero knowledge content extractable. --decide ... --decision discard REFUSED (discard operates on candidate IDs, not sources); source returned to queue via --release cc1760193b07a0c7 → back to pending.
Contradictions found across the batch (mandatory pass): none (single substantive source; homepage discarded as non-substantive).
Pages created (1 entity): vibe-coder-mcp
Existing pages updated: 1 — added [[vibe-coder-mcp]] forward-link to token-optimizer-mcp Related (reciprocal token-cost peer, resolves orphan without backfilling).
Sources completed: c9c40b79772056c9 after lint gate passes.
Tags: prompt-engineering, tool-origin, workflow-pattern, token-usage-reduction — all existing vocabulary; no _schema.md changes.

Gains / Losses

Gained: vibe-coder-mcp entity (MCP-server tool-origin in the token-reduction cluster, reciprocal to token-optimizer-mcp); learned --decide only accepts candidate IDs and --release returns a claimed source to pending without a candidate.
Losses/complexity: +1 forward-link edit on token-optimizer-mcp (now 7 outbound); 1 source still pending (developers-openai-com-9dcdba4f) with no candidate — final disposition (user, 2026-08-01): keep pending in the queue as a permanent non-item; it blocks nothing (backlog is unbounded) and preserves the immutable capture. No --complete will be run for it.

[2026-08-01] maintenance | Rename graphify concept → graphify-skill (cross-wiki slug collision)

Trigger: full-vault lint (scripts/wikilint.ps1 -Scope all, 233 pages) flagged duplicate slug graphify — main-wiki entity wiki/entities/graphify.md (created 2026-07-14) vs prompt-wiki concept created in Batch 10. Prompt-scoped lint passes only check within wiki/prompts/; only -Scope all surfaces cross-wiki collisions.
Root cause: Batch 10 routing decision routed Graphify material to the prompt wiki as a tool-origin concept while a main-wiki Graphify entity already existed, producing a colliding slug.
Fix applied: wiki/prompts/concepts/graphify.md → wiki/prompts/concepts/graphify-skill.md; title updated to “Graphify (skill)”. No aliases set (graphify-skill = slug redundant; graphify would collide with the main-wiki real slug — Check 11).
Inbound links updated: tool-search-tool.md Related ([[graphify]] → [[graphify-skill]]); prompts-index.md Concepts list.
Left untouched: main-wiki knowledge-graph.md and external-links-connections-deep-dive.md links — they resolve to the main-wiki graphify entity, which is the correct target.
Verification: wikilint -Scope all now wikilint: OK (233 pages).

Gains / Losses

Gained: full-vault lint green; duplicate slug eliminated; prompt-wiki Graphify content preserved under a distinct slug.
Losses/complexity: slug graphify-skill is less discoverable than graphify for the prompt wiki; one historical Batch 10 log entry still references the old [[graphify]] slug (append-only logs are not retro-edited).

[2026-08-02] ingest | Claude Opus 5 System Prompt (leaked, consumer chat)

Trigger: user requested capture of the leaked Claude Opus 5 system prompt from the asgeirtj/system_prompts_leaks mirror.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0.576; main 13.8208 / prompts 18.7988), top prompts match grok-personas. Manually overrode to the prompt wiki — the source is a leaked real system prompt, which matches the prompt-wiki trigger.
Capture: https://raw.githubusercontent.com/asgeirtj/system_prompts_leaks/main/Anthropic/claude-opus-5.md → raw/prompts/leaked/claude-opus-5.md. Provenance frontmatter prepended (type: leaked-prompt, created: 2026-08-02, source_type: leaked-prompt, source_url). Body immutable.
Page created: wiki/prompts/entities/claude-opus-5-system-prompt.md — new entity, prompt_type: system, envelope_format: markdown, tool_origin: claude, confidence: low, lifecycle: l1. Documents claude_behavior block, persistent memory filesystem toolset, Fable-safeguard rerouting, safety/wellbeing surfaces (end_conversation, privacy omission), and the # functions JSON tool block.
Linked from existing: claude-code-system-prompt.md gained a sibling-of forward link to resolve the orphan (forward-only, no backfill beyond that sibling link).
Fixed during lint: 2 errors on the new page — orphan (fixed via forward link) and unrecognized relationship type contrasts-with (changed to is-sibling-of grok-4-5-system-prompt); also corrected a stale [[intent-classification]]/[[grok-4.5-system-prompt]] typo references.
Catalog/log: prompts-index.md Entities +1 (56 total); this log entry added.
Verification: wikilint -WikiRoot wiki/prompts → wikilint: OK (see below).

Gains / Losses

Gains: new comparative evidence for consumer- vs agent- Claude prompt design; persistent-memory (filesystem) tool family now captured at consumer level.
Losses/complexity: confidence: low (leaked mirror, not vendor-published); uncertainties logged as Open Questions on the page.

[2026-08-02] ingest | Claude Code Opus 5 System Prompt (leaked, real CLI)

Trigger: user requested capture of the real Claude Code CLI system prompt (Opus 5) from the asgeirtj/system_prompts_leaks mirror.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0.598; main 13.0159 / prompts 19.3906), top prompts match claude-opus-5-system-prompt. User confirmed override to the prompt wiki (leaked real system prompt; classifier’s prompts score dominates).
Capture: https://raw.githubusercontent.com/asgeirtj/system_prompts_leaks/main/Anthropic/Claude%20Code/claude-code-opus-5.md → raw/prompts/leaked/claude-code-opus-5.md (2,519 lines). Provenance frontmatter prepended (type: leaked-prompt, created: 2026-08-02, source_type: leaked-prompt, source_url). Body immutable.
Page created: wiki/prompts/entities/claude-code-opus-5-system-prompt.md — new entity, prompt_type: system, envelope_format: markdown, tool_origin: claude-code, confidence: low, lifecycle: l1. Documents the system surface, per-project memory filesystem (MEMORY.md index, user|feedback|project|reference taxonomy), ~37-tool surface (task/cron/schedule/worktree/plan gates), skill loading, and security tiers.
Linked from existing: claude-code-system-prompt.md gained a sibling-of forward link to the new page (primary-source vs synthesized).
Catalog: prompts-index.md Entities +1 (57 total); this log entry appended.
Verification: wikilint -WikiRoot wiki/prompts → wikilint: OK (below).

Gains / Losses

Gains: first primary-source (real leaked CLI prompt) for Claude Code in the prompt wiki; rich new tool taxonomy (cron/task/worktree/monitor) and a second memory-layer model.
Losses/complexity: confidence: low (leaked mirror); two entity pages now cover the same CLI (synthesized vs primary-source) — explicitly linked sibling-of, no content intersection removed.

[2026-08-02] ingest | Anthropic Research Instructions (leaked system prompt)

Trigger: user provided the URL to the Anthropic research_instructions.md system prompt from the asgeirtj/system_prompts_leaks mirror for capture.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0.57; main 8.0151 / prompts 10.611), top prompts match Grok 4.5 System Prompt. Manually overrode to the prompt wiki — the source is a leaked Anthropic system prompt configuring “advanced research mode,” which matches the prompt-wiki trigger (real system prompts, leaked). Classifier returned ambiguous because the query text is new content not yet indexed; the prompts score dominates (10.611 > 8.0151) and the content type is unambiguous.
Capture: https://github.com/asgeirtj/system_prompts_leaks/blob/main/Anthropic/research_instructions.md → raw/prompts/leaked/anthropic-research-instructions.md. Provenance frontmatter prepended (title, type: system-prompt, source_url, source_type: web, created: 2026-08-02). Body immutable.
Pages created (3):
- wiki/prompts/entities/anthropic-research-instructions.md — entity, prompt_type: system, envelope_format: xml, tool_origin: claude, confidence: low, lifecycle: l1. Documents the tool-priority override, clarifying-questions protocol, copyright/safety guardrails, and envelope/tool schema.
- wiki/prompts/concepts/clarifying-questions.md — concept, confidence: medium, lifecycle: l1. The ≤3-questions-then-launch pattern with worked examples and comparison to other systems.
- wiki/prompts/concepts/tool-priority-override.md — concept, confidence: medium, lifecycle: l1. The “one tool takes absolute priority” architectural pattern.
Pages updated (extended, 0 existing — all new): none extended; no contradictions with existing content.
Navigation updates: prompts-index.md — Pages 137 → 140 (Entities 57→58, Concepts 69→71). Both new concepts added to Concepts list alphabetically; new entity added to top of Entities list.
Verification: wikilint -Scope prompts → wikilint: OK — All checks passed (140 pages, scope: prompts). LOW CONF advisories are expected (leaked-prompt pages carry confidence: low by design).
Route validation: classifier returned ambiguous (confidence: 0.57; main 8.0151 / prompts 10.611), top prompts match Grok 4.5 System Prompt. Manually overrode to prompt wiki — content is a leaked Anthropic system prompt, the prompt wiki’s core domain. Prompts score dominates (10.611 > 8.0151); classifier ambiguity is due to new content not yet indexed.

Gains / Losses

Gains: captured first Anthropic “advanced research mode” system prompt in the vault; introduced two new architectural concepts (tool-priority-override, clarifying-questions protocol) that generalize beyond this single prompt.
Losses/complexity: +2 concept pages and +1 entity page add surface area; the tool-priority-override pattern overlaps conceptually with tool-context-management (explicitly marked as differing). confidence: low on the entity reflects the leaked-mirror provenance.

[2026-08-02] ingest | Gemini Deep Research Prompt (shared gist)

Trigger: user provided the URL to a shared Gemini Deep Research system prompt gist (swayson/.../gemini-deep-research-prompt) for capture.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0; main 0 / prompts 0, no trained match). User confirmed manual override to the prompt wiki — the source is a real shared system prompt for a research tool, which matches the prompt-wiki trigger (“real system prompts, leaked”). Classifier ambiguity is expected for new gist content with no indexed match.
Capture: https://gist.githubusercontent.com/swayson/ed7d8db74dd9c50a3b5568cd969ccad8/raw/.../gemini-deep-research-prompt → raw/prompts/leaked/gemini-deep-research-prompt.md. Provenance frontmatter prepended (title, type: system-prompt, source_url, source_type: web, created: 2026-08-02). Body immutable.
Pages created (2):
- wiki/prompts/entities/gemini-deep-research-prompt.md — entity, prompt_type: system, envelope_format: markdown, tool_origin: gemini, confidence: low, lifecycle: l1. Documents the five-phase TDP/MIG/CAS/RG/IR workflow, artifacts-as-output, three-stage bias mitigation, style rules, and contrasts with Anthropic/Perplexity research prompts.
- wiki/prompts/concepts/documented-research-process.md — concept, confidence: high, lifecycle: l1. The process-over-product research workflow pattern: mandated ordered phases whose artifacts (source notes tables, evaluation matrices, iteration logs) are part of the deliverable.
Pages updated (2, within the 5-page batch cap):
- wiki/prompts/entities/gemini.md — added gemini-deep-research-prompt forward link (orphan resolution, forward-only); updated: 2026-08-02.
- wiki/prompts/concepts/prompt-engineering-guides.md — OpenAI Deep Research section cross-references gemini-deep-research-prompt and documented-research-process; sources += raw path; updated: 2026-08-02.
Contradictions: none — positioned as complement/contrast to anthropic-research-instructions (tool-gating vs process-gating) and two-tier-search-architecture (visible vs invisible planning); no contested flags.
Gap noted: raw/prompts/leaked/openai-tool-deep-research.md (captured 2026-08-02) still has no wiki page — out of scope for this ingest, flagged for a future pass.
Navigation updates: prompts-index.md — Pages 140 → 142 (Entities 58→59, Concepts 71→72; also corrected stale Concepts section header 69→72). This log entry appended.
Verification: wikilint -Scope prompts → see output below. LOW CONF advisory expected on the entity (gist provenance, not vendor-authenticated).

Gains / Losses

Gains: first Gemini research-mode prompt in the vault; new documented-research-process concept generalizes the process-over-product pattern beyond this single gist and contrasts cleanly with the two existing research-prompt architectures.
Losses/complexity: confidence: low on the entity (third-party gist provenance); the prompt’s heavy process overhead is noted as an Open Question (verbosity/output-token cost vs rigor).

[2026-08-02] ingest | How to Reduce MCP Token Costs for Claude Code at Scale (maxim.ai)

Trigger: user shared the maxim.ai article URL for capture.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0; main 0 / prompts 0, no trained match). User confirmed manual override to the prompt wiki — the article targets Claude Code MCP token costs, matching the token-usage-reduction / tool-context-management cluster. Classifier ambiguity is expected for new content with no indexed match.
Capture: https://www.getmaxim.ai/articles/how-to-reduce-mcp-token-costs-for-claude-code-at-scale/ → raw/prompts/articles/bifrost-mcp-gateway-claude-code-token-costs.md. Provenance frontmatter prepended (title, type: article, source_url, source_type: web, created: 2026-08-02). Body immutable.
Pages created (2):
- wiki/prompts/entities/bifrost.md — entity, confidence: medium, lifecycle: l1. Maxim AI’s open-source AI gateway; Code Mode (MCP tool surfaces as Python stubs + four meta-tools + sandboxed Starlark), per-tool virtual keys, MCP Tool Groups, single /mcp endpoint, observability. Vendor benchmarks (58/84/92% input-token cuts at 96/251/508 tools; 11µs @ 5k rps) marked confidence: low.
- wiki/prompts/concepts/code-mode-orchestration.md — concept, confidence: medium, lifecycle: l1. The generalized gateway pattern: cost bounded by what the model reads, not tool count; corroborated by Anthropic code-execution-with-mcp (~150k→2k tokens) and Cloudflare’s Code Mode blog.
Pages updated (2, within the 5-page batch cap):
- wiki/prompts/concepts/tool-context-management.md — added “The Gateway Layer” section (fifth, infrastructure-level control: gateway decides what the client can even see) + code-mode-orchestration link; sources += raw path; updated: 2026-08-02.
- wiki/prompts/concepts/token-usage-reduction.md — added MCP-overhead evidence (schemas re-enter context every message; ~7k/msg typical, 15–20k/turn heavier, 50k+ extremes) + the gateway as fleet-level lever; sources += raw path; updated: 2026-08-02.
Contradictions: none flagged. Article’s cited third-party numbers (jdhodges ~7k/msg, mindstudio 15–20k/turn) are references, not captured evidence (Guardrail #16); referenced in prose only. Cross-linked token-reduction-is-not-cost-reduction (input cuts ≠ billed cost) as an Open Question rather than settled equivalence.
Navigation updates: prompts-index.md — Pages 142 → 144 (Entities 59→60, Concepts 72→73). This log entry appended.
Verification: wikilint -Scope prompts → see output below.

Gains / Losses

Gains: first MCP-gateway entity in the vault; new code-mode-orchestration concept generalizes the stub-filesystem + sandboxed-script pattern beyond the vendor; the MCP-overhead slice adds a distinct input-spend lever to token-usage-reduction.
Losses/complexity: vendor benchmarks are self-published (confidence: low) — claims need independent measurement before treated as settled; code-mode and tool-context-management programmatic tool calling overlap conceptually (explicitly cross-referenced as complement, not duplicate).

[2026-08-02] ingest | Reduce Token Waste (gomzkov, Medium)

Trigger: user shared the Medium article URL (https://gomzkov.medium.com/reduce-wasted-tokens-d36eb5b46442) for capture.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0; main 0 / prompts 0, no trained match). User confirmed manual override to the prompt wiki — all four named tools (RTK, Headroom, Ponytail, Caveman) are already prompt-wiki entities; the article is squarely the token-usage-reduction / tool-origin domain. Classifier ambiguity expected for new content with no indexed match.
Capture: raw/prompts/articles/reduce-wasted-tokens.md — provenance frontmatter prepended (title, type: article, source_url, source_type: web, created: 2026-08-02). Body immutable, faithful to the fetched article (six waste sources, five tool sections, “BUT” section, sources list).
Pages created (2):
- wiki/prompts/entities/codegraph.md — entity, confidence: low, lifecycle: l1. Repo-navigation map for coding agents (symbols/callers/dependencies); vendor benchmark 58% fewer tool calls, 22% faster, file reads ~zero across 7 repos. Distinct from the vault’s codebase-memory (tree-sitter graph over MCP) — cross-linked as adjacent mechanism.
- wiki/prompts/concepts/token-waste-sources.md — concept, confidence: medium, lifecycle: l1. The diagnostic framing: “where is my agent actually wasting context?” with six waste sources mapped to fix classes + the measurement checklist (input/output/reasoning tokens, tool calls, latency, cost per task, tests passed, reruns, context-loss frequency).
Pages updated (3, within the 5-page batch cap):
- wiki/prompts/concepts/token-usage-reduction.md — added waste-source-diagnosis and repo-navigation bullets + token-waste-sources / codegraph links; sources += raw path; updated: 2026-08-02.
- wiki/prompts/entities/caveman.md — added “Caveats” section: third-party coding benchmark measured a 7% increase in total tokens (output is the cheap part; reasoning tokens untouched); sources += raw path; updated: 2026-08-02.
- wiki/prompts/entities/headroom.md — added survey fact that Headroom uses RTK under the hood for shell output (not two fully separate layers) + independent 47–92% savings range; sources += raw path; updated: 2026-08-02.
Contradictions: none new — the article supports the existing contested stance on input-compression proxies (RTK issue #582, token-reduction-is-not-cost-reduction). The Caveman 7% counterexample was added as a caveat, not a contradiction, since it confirms rather than conflicts with the asymmetry already documented. “Headroom uses RTK under the hood” recorded as a new fact on the Headroom page.
Navigation updates: prompts-index.md — Pages 144 → 146 (Entities 60→61, Concepts 73→74). This log entry appended.
Verification: wikilint -Scope prompts → see output below.

Gains / Losses

Gains: first repo-navigation entity (codegraph) and the waste-source diagnostic concept; the six-source taxonomy gives a principled way to choose among the token-reduction tools the vault already tracks; the Caveman counterexample strengthens the input/output asymmetry evidence base.
Losses/complexity: codegraph benchmark is vendor-reported (confidence: low); token-waste-sources overlaps with the tool-decision surface already scattered across tool entities (explicitly cross-linked, not duplicated).

[2026-08-02] ingest | Claude Code MCP Servers and Token Overhead (MindStudio)

Trigger: user shared the MindStudio blog URL (https://www.mindstudio.ai/blog/claude-code-mcp-server-token-overhead) for capture.
Routing: scripts/route-classifier.ps1 returned ambiguous (confidence: 0; main 0 / prompts 0, no trained match). User confirmed manual override to the prompt wiki — the article is the primary source behind the 15–20k/turn MCP-overhead figure the Bifrost ingest already cited; squarely the token-usage-reduction / tool-context-management cluster. Classifier ambiguity expected for new content with no indexed match.
Capture: raw/prompts/articles/mindstudio-claude-code-mcp-token-overhead.md — provenance frontmatter prepended (title, type: article, source_url, source_type: web, created: 2026-08-02). Body immutable, faithful to the fetched article (cost model, 18k breakdown, 4-step audit, reduction strategies, MindStudio Agent Skills Plugin section, FAQ, key takeaways).
Pages created (1):
- wiki/prompts/concepts/capability-as-method-calls.md — concept, confidence: medium, lifecycle: l1. The third architectural answer to tool-schema bloat: capabilities exposed as typed method calls in agent code (agent.searchGoogle()) with execution (rate limiting, retries, auth) outside the prompt — distinct from client-side deferral (tool-context-management) and gateway orchestration (code-mode-orchestration).
Pages updated (2, within the 5-page batch cap):
- wiki/prompts/concepts/token-usage-reduction.md — added the primary-source cost model (100–500 tokens/tool; 1,500–4,000/10-15 tools; 5,000–8,000/30-tool server; 12–20k multi-server), the explicit no-cross-turn-caching FAQ, and the 4-step audit process + pruning/project-config/description-trimming/filtering levers; sources += raw path; updated: 2026-08-02.
- wiki/prompts/concepts/tool-context-management.md — added “Measuring and Trimming Schema Cost” section (audit formula (tools × 200) + (desc_chars ÷ 4), client-side trims: prune, project-level configs, source description trimming, tool filtering, per-task agents; session-init overhead); sources += raw path; updated: 2026-08-02.
Contradictions: none flagged. The article corroborates existing content — the 18k mid-range figure independently supports the Bifrost-cited 15–20k/turn range, and the FAQ explicitly confirms “no cross-turn caching of tool definitions at the API level,” which resolves the mechanism question the gateway evidence raised. Vendor promotion of its own Agent Skills Plugin in §5 treated as directional, not benchmarked.
Navigation updates: prompts-index.md — Pages 146 → 147 (Concepts 74→75). This log entry appended.
Verification: wikilint -Scope prompts → see output below.

Gains / Losses

Gains: primary source behind the previously reference-only mindstudio 15–20k/turn figure; concrete audit methodology (formula + 4 steps) that fills the diagnostic half of token-waste-sources; a genuinely distinct third architecture (capability-as-method-calls) for schema bloat; explicit no-caching confirmation resolves the gateway-vs-cache mechanism question.
Losses/complexity: cost model and reduction claims are vendor-authored (confidence: medium on the concept, estimates not measurements); capability-as-method-calls rests on a single vendor-promo section (its scalability and correctness-at-scale are recorded as Open Questions).

[2026-08-03] query | System prompts — prompt-engineering side and leaked/real captures

Route: route-classifier.ps1 -Query "What do we know about system prompts - prompt-engineering side and real/leaked system prompts captured" returned ambiguous (confidence 0.526, main 17.17 “Context Engineering” vs prompts 19.06 “Grok Expert”). User explicitly asked not to be prompted to pick a wiki; both catalogs consulted rather than pausing.
Catalog query: Found 10 leaked/real system-prompt raw captures in raw/prompts/leaked/ (anthropic-research-instructions, claude-code-opus-5, claude-opus-5, gemini-deep-research-prompt, grok-4.5-system-prompt, grok-build, grok-expert, grok-personas, openai-tool-deep-research, perplexity-ai-system-prompt). Analyzed 3 in depth: how-to-structure-a-system-prompt (five-universal-sections taxonomy, ~52-prompt corpus), claude-code-system-prompt (markdown envelope, task-state-machine, cooldown-heuristics), claude-opus-5-system-prompt (leaked, confidence low — public mirror, not vendor-authenticated; claude_behavior block, memory filesystem, Fable-safeguard rerouting). Remainder summarized from prompts-index.md one-liners only.
Verification: drafted answer saved to a temp file and run through scripts/query-check.ps1 -Scope prompts -Catalog; first run FAILED on route-header and catalog-counts formatting (missing literal Route: prompts line and Found:/Analyzed: labels); answer text corrected and rerun printed query-check: OK (all 9 checks PASS).
Action: read-only synthesis, answered directly in chat; not filed as a new query page (not requested).

[2026-08-03] ingest | JetBrains: Ponytail Skill for Claude Code — Does It Really Cut Agent Code by 54%?

Trigger: user shared the JetBrains AI blog URL (https://blog.jetbrains.com/ai/2026/07/ponytail-skill-claude-tested/) for capture.
Routing: scripts/route-classifier.ps1 -Query "ingest jetbrains blog post about ponytail skill tested with claude" returned ambiguous (confidence: 0; main 0 / prompts 0, no trained match). User confirmed manual override to the prompt wiki — the article is Part 3 of JetBrains’ “token saver” series on public skills the vault already tracks as prompt entities (ponytail, caveman, rtk-rust-token-killer); independent measured data about real agent skills is squarely the token-usage-reduction / skill-verification cluster. Initial manual read suggested main wiki; corrected to prompt wiki once the existing wiki/prompts/entities/ponytail.md was found.
Capture: raw/external/blog-jetbrains-com-ponytail-skill-claude-tested-b56c8490.md via scripts/external_links.py capture — HTTP 200, content_hash: b56c84903de77a6a, queued for review (id a14d7c3bb6565b87). Immutable snapshot with provenance frontmatter (title, type: web-snapshot, source_url, source_type: web, http_status, content_hash).
Pages updated (3, within the 5-page batch cap; none created):
- wiki/prompts/entities/ponytail.md — added “Independent Measurement (JetBrains, 2026-08-03)” section: 80 paired SkillsBench tasks, claude-sonnet-5 medium, verifier-scored → measured −15% code / −10.3% cost (p=0.004) / −11% time vs advertised −54%/−22%/−27%; first statistically solid cost saving in the series; self-activation zero times across 10 sessions (plain SKILL.md install measures nothing; plugin SessionStart hook is the only effective path); cut concentrates on over-build traps (−31% large builds, ~0 on lean); quality null (65 identical/9 worse/6 better — verifiers don’t test security/validation/accessibility so the “100% safe” claim is untested); adoption audited 100%/0%; small samples lie (10-task smoke said +9.6% cost). Open Questions += workload-shape of the −54% gap. Related += caveman. sources += raw path; updated: 2026-08-03.
- wiki/prompts/entities/caveman.md — Caveats section += Part 1 independent measurement: −8.5% code vs advertised −65%, ~10% cost only after removing a pricing-tier outlier (“fragile number”); consistent with the earlier 7% total-token-increase report; the run tested the skill only, not the caveman wrap proxy or CaveGemma. Related += ponytail; sources += raw path; updated: 2026-08-03.
- wiki/prompts/entities/rtk-rust-token-killer.md — Contradiction section += Part 2 independent measurement: +7.6% cost vs advertised −60–90%, “every bit as significant, just pointing the wrong way” — independent corroboration of issue #582’s 18% cost-increase mechanism from a completely different direction. Open Questions += workload sensitivity. Related += ponytail; sources += raw path; updated: 2026-08-03.
Contradictions: the JetBrains measurements support the existing contested stance on input-compression proxies (rtk-rust-token-killer contested: true; token-reduction-is-not-cost-reduction) and the existing caveat that vendor token-claims must be re-measured. Vendor-vs-measured gaps on all three skills were recorded as dated positions on each entity page, not as page-level contradictions. No new page-level contested flags.
Navigation updates: prompts-index.md — page count unchanged (147; 3 entities updated, 0 created). One-liners refreshed for ponytail, caveman, rtk-rust-token-killer with measured figures. This log entry appended.
Verification: wikilint -Scope prompts → see output below.

Gains / Losses

Gains: first independent third-party measurement for three of the vault’s token-reduction entities, all from one methodologically candid source (paired runs, pinned commits, audited adoption, self-reported contamination bug); resolves the direction of the ponytail claim (−54% is workload-shaped, not marketing-shaped) and gives the strongest negative evidence yet for token-reduction-is-not-cost-reduction.
Losses/complexity: measured figures come from a single vendor-agnostic blog (JetBrains), itself bounded by SkillsBench’s task mix (conservative for the code claim, so the −54% is neither confirmed nor refuted); no new page created — the material is an extension of existing entities rather than a new concept.

[2026-08-03] lint | synaptic-lint full-vault health check

Scope: all. Structural command: pwsh -File scripts/wikilint.ps1 -Scope all -> wikilint: OK (256 pages): Property 0 errors, Link 0 errors, Content 0 errors. Advisory layer reported 15 low-confidence warnings across 11 prompt pages; no stale pages, provenance-floor failures, or broken external links.
Prompt semantic audit (147 active pages; 141 prompt raw sources; 56 provenance-valid external snapshots): no direct active-page citations to main-only raw roots and no unrecorded contradiction confirmed. raw/prompts/leaked/openai-tool-deep-research.md remains substantive evidence without a contributing page. The duplicated taxonomy-synthesis sources trace to a main-only raw source and underpin 41 prompt pages, creating a route-provenance dependency. chain-of-thought/chain-of-thought-pattern remains a distinction-versus-merge review candidate.
Main semantic findings are recorded in log.md. No raw evidence or wiki page was modified during this read-only audit.

Gains / Losses

Gains: structural health and the unrepresented OpenAI research-prompt source are recorded without conflating semantic review with the structural pass.
Losses/complexity: the taxonomy provenance dependency and review candidates remain open; this audit added log history only.

[2026-08-04] ingest | Queue completion: ponytail JetBrains capture (a14d7c3bb6565b87)

Trigger: main-wiki backlog batch surfaced queue id a14d7c3bb6565b87 (raw/external/blog-jetbrains-com-ponytail-skill-claude-tested-b56c8490.md) still pending.
Routing: re-run of scripts/route-classifier.ps1 returned ambiguous (confidence: 0); user confirmed the prompt wiki (already the home of the three token-saver entities).
Finding: the snapshot was already fully ingested on 2026-08-03 — ponytail.md, caveman.md, and rtk-rust-token-killer.md all cite the raw path and carry the measured figures; the queue item was simply never marked complete after that ingest. No page changes were needed; the compounding invariant required no re-edit.
Action: python scripts/auto_ingest.py --complete a14d7c3bb6565b87; entry recorded in the main-wiki log.md batch entry. No prompt-wiki page touched; wikilint -Scope prompts → wikilint: OK — All checks passed (147 pages, scope: prompts).
Gains: queue backlog cleared of a duplicate-consume case without duplicating content or fabricating a page. Losses: none for the prompt wiki (log history only).

[2026-08-06] ingest | Hugging Face Advanced Prompt Engineering survey

Source: raw/external/huggingface-co-advanced-prompt-engineering-bc976d6b.md (captured 2026-08-06, http 200). Copied from the evaluation worktree raw/external/ into the main vault so it is available for cross-wiki ingest.
Routing: scripts/route-classifier.ps1 -Query "...advanced prompt engineering techniques" → ambiguous (confidence: 0.577), main 6.80 / prompts 9.30, top match Gemini Advanced. The PowerShell wrapper’s fallback returns ambiguous (0.0) for all queries on WSL (helper path resolved to nil), so the direct Python helper output was used: prompts score > main → prompt wiki route (confirmed by article subject = prompt-engineering technique survey).
Pages created (2):
- constraint-based-prompting (concept, l0) — output-format/reasoning/behavioral constraint prompting.
- persona-engineering (concept, l0) — role-based prompting / persona assignment.
Pages extended (5): chain-of-thought (→l1), self-consistency (→l1), rag-prompting (→l1), llm-settings-and-parameters (→l2), structured-outputs (→l2). All extended not overwritten; new source added to each sources.
Catalog/log: prompts-index.md — Concepts 75→77, total 147→149 content. This prompts-log.md entry appended.
Contradictions: none; no cross-source conflict with this survey (consistent restatement of existing PE knowledge, no contradicting positions recorded).
Gains: two new technique concepts fill genuine gaps; five rich pages gained a correlated source and upgraded lifecycle. Losses: none to content; note that route-classifier.ps1 on WSL silently masks Python failures as ambiguous (a tooling defect to fix separately).
Verification: run pwsh -File scripts/wikilint.ps1 -Scope prompts → see appended result line.

LLM Wiki — Synaptic Lattice

Explorador

prompts-log.md — Prompt Wiki Chronological Action Log

[2026-08-07] ingest | Prompting Guide techniques index (backlog cleanup)

Source

Route validation

Pages

Discarded (not ingested — no wiki coverage needed)

Verification

[2026-08-07] query | Polimento de prompt de pesquisa (Superpowers × Matt Pocock)

[2026-08-07] query | Craft prompt to analyze and prioritize a 35-skill package

[2026-08-06] ingest | Tree of Thoughts technique from Prompt Engineering Guide

Source captured

Route validation

Pages extended

Navigation

Verification

[2026-08-06] ingest | ReAct Prompting technique extension

Source

Route validation

Pages extended

Verification

[2026-08-06] ingest | Graphify and Capacities updates from external snapshots

Sources recaptured

Pages extended

Verification

[2026-08-06] ingest | nashsu/llm_wiki and claude-obsidian updates

Sources recaptured

Pages extended

Verification

[2026-08-06] ingest | Tree of Thoughts (ToT) technique from promptingguide.ai

Source

Route validation

Pages

Verification

[2026-08-06] ingest | OpenAI Prompt Caching Guide (detailed)

Source

Route validation

Pages updated (3 content pages, within 5-page batch cap)

Navigation updates (2)

Guardrails applied

Verification

[2026-08-03] audit | Prompt Wiki content maintenance closed

prompts-log.md — Chronological Action Log

[2026-07-16] create | Prompt wiki initialized

[2026-07-17] ingest | Prompt Engineering Guide (promptingguide.ai/pt)

[2026-07-22] index | Concept catalog sync (28→44)

[2026-07-22] index | Comparison catalog sync (0→5)

[2026-08-03] query | Vault coverage of system prompts (theory + real captures)

[2026-07-22] audit | Remaining raw source audit

[2026-07-22] lint | Final prompt wiki lint

[2026-07-22] query | Voice/style prompting techniques

[2026-07-31] ingest | Grok Personas leaked system-prompt collection

Source

Route validation

Pages created (1 entity)

Pages updated (1 entity)

Pages updated (2 navigation)

Wikilinks

Guardrails applied

Verification

[2026-07-31] ingest | Grok Expert multi-agent team-leader system prompt

Source

Route validation

Pages created (1 entity)

Pages updated (2 entities)

Pages updated (2 navigation)

Wikilinks

Guardrails applied

Verification

[2026-07-31] query | Audit and enhance token-reduction catalogue prompt

[2026-07-31] ingest | Grok Build agentic coding CLI system prompt

Source

Route validation

Pages created (1 entity, 1 concept)

Pages updated (3 entities)

Pages updated (2 navigation)

Wikilinks

Guardrails applied

Open gap (noted, not created)