AGENTS.md — Prompt Engineering Wiki

Schema and operations for the prompt engineering sub-wiki. Consult this wiki when the user needs to create, improve, evaluate, or understand prompts. Pipeline operations reference the root AGENTS.md (this file only covers prompt-specific schema and taxonomy).

Domain

Prompt engineering wiki. Consulted to create, improve, evaluate, or understand prompts. Includes real system prompts (leaked), product configs, technique articles, and case studies.

Frontmatter Schema

Base fields (OKF-compliant)

title: Page Title
type: entity | concept | comparison | query
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [tag1, tag2, tag3]
sources: [raw/prompts/articles/source-name.md]
confidence: high | medium | low
lifecycle: l0 | l1 | l2 | l3
contested: false
contradictions: []

Prompt-specific fields

Field	When used	Values	Required?
prompt_type	entity of real prompt	system, user, tool-schema, product-config	Yes for prompt entities
envelope_format	entity of real prompt	xml, markdown, yaml, json, hybrid, template-vars	Yes for prompt entities
tool_origin	entity of real prompt	claude-code, cursor, windsurf, v0, cline, manus, perplexity, etc.	Yes for prompt entities
patterns_used	entity or concept	[chain-of-thought, task-state-machine, plan-gate, ...]	Optional
token_count_estimate	entity of real prompt	integer (e.g. 30000)	Optional

Accepted Page Types

• entity: Real prompt (system, user, tool-schema, product-config) — requires prompt_type, envelope_format, tool_origin
• concept: Technique or pattern (chain-of-thought, xml-envelope, anti-injection, etc.)
• comparison: Side-by-side analysis (xml-vs-markdown-envelopes, monolithic-vs-multi-file)
• query: Filed answer (how-to-structure-a-system-prompt)

Operations

Query

Follow the pipeline defined in root AGENTS.md (Query section), scoped:

• A2 (index.md) → prompts-index.md
• A1 (grep) → raw/prompts/ + wiki/prompts/
• A3 (read) → pages in wiki/prompts/

Ingest

Follow the pipeline defined in root AGENTS.md (Ingest section, 9 steps), scoped:

• raw/prompts/* instead of raw/*; captured external snapshots in raw/external/ are also valid evidence when their capture metadata is present and the source is prompt-domain material
• prompts-index.md instead of index.md
• prompts-log.md instead of log.md

Lint

Run scripts/wikilint.ps1 with -WikiRoot pointing to wiki/prompts/:

.\scripts\wikilint.ps1 -WikiRoot "$PSScriptRoot\..\wiki\prompts"

Tag Taxonomy

• prompt-engineering — Root tag for all prompt wiki pages
• system-prompt — Full system prompt (not just a fragment)
• envelope-format — XML, markdown, YAML, JSON envelope
• workflow-pattern — Task state machine, plan gate, cascade
• safety-defense — Anti-prompt-injection, anti-spill
• communication-style — Solution-only, explain-before-do, paired
• architecture-pattern — Monolithic, multi-file, SDK
• tool-origin — Claude Code, Cursor, Windsurf, etc.
• instruction-pattern — CoT, few-shot, ReAct
• research-gap — Identified gap in current practice

Guardrails

• Do not accept creative prompts (“write a poem”) as entity
• Do not overwrite real prompts captured in raw/prompts/leaked/
• Max 8 outbound links per page
• Forward-only linking (never backfill old pages)
• 2+ outbound wikilinks per page (or cross-wiki links to main wiki pages)
• All tags must be in taxonomy above

Key concepts: xml-envelope-format, chain-of-thought-pattern, anti-prompt-injection, communication-style-spectrum

Routing Back to Main Wiki

If a query involves fine-tuning, RAG pipelines, second brain architecture, OKF specification, or LLM architecture → route to root AGENTS.md (main wiki).

Cross-wiki links use double-bracket syntax with a wiki/ prefix (code-escaped in this doc) to reference main wiki pages. The main wiki’s page about the LLM Wiki pattern is the most common cross-reference.

Sources

• [docs/archive/2026-07-16-prompt-engineering-wiki-design.md]