LLM Wiki — Synaptic Lattice

prompts-index.md — Prompt Engineering Wiki Catalog

Properties1
tagsnavigation, prompt-engineering

Leitura de 12 min

Prompt Engineering Wiki

Sub-wiki catalog. Updated: 2026-08-09. Pages: 150 content (61 entities, 78 concepts, 5 comparisons, 6 queries)

Entities (61)

  • anthropic-research-instructions — Anthropic Research Instructions (leaked, system prompt) — advanced research mode overlay; tool-priority override, clarifying-questions protocol, copyright/safety guardrails

  • claude-code-opus-5-system-prompt — Claude Code (Opus 5) System Prompt (leaked, real CLI) — primary-source harness prompt; memory filesystem, task/cron/schedule tools, security tiers

  • claude-opus-5-system-prompt — Claude Opus 5 System Prompt (leaked, consumer chat) — claude_behavior block, persistent memory filesystem, Fable-safeguard rerouting

  • caveman — token-efficient agent stack (skill + engine + enterprise): /caveman skill cuts ~65% output tokens, caveman wrap byte-safe proxy, CaveGemma fine-tune, install across 30+ agents; JetBrains measured −8.5% code vs advertised −65%

  • opencode — OpenCode (Anomaly) — terminal-native agentic coding CLI, V2; whitelist permissions, checkpoint compaction, naming collision with archived Go project

  • vibe-coder-mcp — MCP server extending Cursor/Cline/Claude Desktop with 10 agentic tools (RDD task manager, PRD/user-story generators); OPENROUTER_API_KEY required; README-claimed 95-97% token reduction

  • grok-4-5-system-prompt — análise cautelosa do prompt vazado atribuído ao Grok 4.5

  • perplexity-system-prompt — Perplexity AI’s search assistant system prompt (leaked)

  • perplexity-ai — AI-powered search engine and answer engine

  • dair-ai — DAIR.AI — organization behind the Prompt Engineering Guide

  • flan — FLAN (Google) — instruction-tuned T5/PaLM, 1.8K tasks

  • chatgpt — ChatGPT (OpenAI) — GPT-3.5-turbo conversational RLHF model

  • llama — LLaMA (Meta) — efficient foundation model family 7B-65B

  • gpt-4 — GPT-4 (OpenAI) — multimodal text+image, top-10% bar exam

  • gpt-5-6 — GPT-5.6 (OpenAI) — reasoning model family: effort/mode/context params, reasoning-token budget mechanics

  • mistral-7b — Mistral 7B (Mistral AI) — efficient 7B model

  • gemini — Gemini (Google) — multimodal Ultra/Pro/Nano family

  • gemini-advanced — Gemini Advanced (Google) — extended capabilities

  • gemini-deep-research-prompt — Gemini Deep Research Prompt (shared gist) — tool-agnostic, process-centric research prompt; five-phase TDP/MIG/CAS/RG/IR workflow with artifacts-as-output

  • gemini-pro — Gemini 1.5 Pro (Google) — extended context window

  • phi-2 — Phi-2 (Microsoft) — small language model

  • mixtral — Mixtral 8x7B (Mistral AI) — mixture of experts

  • code-llama — Code Llama (Meta) — code-specialized LLaMA variant

  • olmo — OLMo (AI2) — open-source language model

  • sora — Sora (OpenAI) — video generation model

  • claude-3 — Claude 3 (Anthropic) — Haiku/Sonnet/Opus, 200K context

  • gemma — Gemma (Google DeepMind) — open 2B/7B models

  • grok-1 — Grok-1 (xAI) — 314B MoE, Apache 2.0

  • grok-personas — Grok Personas — six leaked system-prompt personas (Companion, Unhinged Comedian, Loyal Friend, Homework Helper, Not a Doctor, Not a Therapist)

  • grok-expert — Grok Expert (Multi-Agent Team Leader) — leaked system prompt for Grok collaborating with a team of agents

  • kimi-k2.5 — Kimi K2.5 (Moonshot AI) — 1T-param MoE, 96.1% AIME

  • llama-3 — Llama 3 (Meta) — 8B/70B/400B, 128K vocab

  • mistral-large — Mistral Large (Mistral AI) — 81.2% MMLU, 32K context

  • mixtral-8x22b — Mixtral 8x22B (Mistral AI) — 141B MoE, 64K context

  • model-collection — Model Collection — 47+ LLMs from BERT to PaLM 2

Tool System Prompts (13)

  • anthropic-tool-use-api — Anthropic Tool Use API — structured API for function-calling and tool integration
  • antigravity-paired-response — Antigravity Paired Response (Google) — dual-response system for paired planning+execution
  • augments-dual-model-prompt — Augment Code Dual-Model Prompt — multi-model architecture for code generation
  • claude-code-system-prompt — Claude Code System Prompt (Anthropic) — terminal-native agentic coding assistant
  • cline-system-prompt — Cline System Prompt — VS Code autonomous coding agent with full XML envelope
  • codebuddy-anti-spill — CodeBuddy Anti-Spill — defense mechanism against prompt extraction for coding assistants
  • cursor-system-prompt — Cursor System Prompt (Cursor AI) — IDE-integrated chat+agent architecture
  • kiro-intent-classifier — Kiro Intent Classifier — lightweight classifier for routing user intent to specialized agents
  • lumo-system-bio — Lumo System Bio (Lumo Labs) — agentic system prompt with biological-metaphor architecture
  • manus-modules-system — Manus Modules System — agentic framework with modular tool orchestration
  • roocode-5-mode-architecture — RooCode 5-Mode Architecture — multi-mode coding agent with specialized profiles
  • windsurf-cascade-system — Windsurf Cascade System — Codeium’s agentic AI coding assistant with cascade workflow
  • xcode-actions-prompt — Xcode Actions Prompt (Apple) — IDE-integrated prompt for Xcode document/explain/playground features
  • grok-build — Grok Build (Agentic Coding CLI) — xAI’s terminal-native coding assistant on Grok 4.5; 27-tool suite, action-safety tiering, Rhai workflows
  • rtk-rust-token-killer — RTK (Rust Token Killer) — CLI proxy compressing command output before context; contested: hook can raise costs 18% (issue #582); JetBrains measured +7.6% cost vs advertised −60–90%
  • headroom — Headroom — context compression layer for AI agents; library/proxy/MCP with output-token shaping
  • bifrost — Bifrost (Maxim AI) — open-source AI gateway; MCP gateway with Code Mode stub filesystem, sandboxed Starlark orchestration, per-tool virtual keys, Tool Groups; vendor-claimed 58–92% input-token reduction
  • codegraph — CodeGraph (colbymchenry) — repo-navigation map for coding agents: symbols/callers/dependencies; 58% fewer tool calls, 22% faster, file reads ~zero (7-repo vendor benchmark)
  • token-optimizer-mcp — Token Optimizer MCP — MCP + plugin optimizing tokens via caching, compression, knowledge-graph feedback; 95%+ claim contested
  • claude-code-cache-fix — claude-code-cache-fix — local proxy repairing silent prompt-cache breaks on /resume; cache-hit observability; “up to 20x cost” on resumed sessions
  • codeburn — codeburn — local-first token/cost tracker across 36 AI coding tools; reads on-disk sessions, never intercepts; optimize/guard/compare subcommands
  • context-mode — Context Mode — cross-platform context-window optimizer; sandboxes tool output (~98% claimed), FTS5 session memory across compaction, hook-enforced routing (18 platforms)
  • crush — Crush (charmbracelet) — terminal-native multi-model agentic coding CLI; MCP + Agent Skills, crushrc bash-config, ~30 providers, FSL-1.1-MIT license
  • ponytail — Ponytail (DietrichGebert) — anti-overbuild skill: YAGNI ladder before writing code; −54% LOC / −22% tokens / 100% safe (agentic benchmark); portable to 20+ agents; JetBrains 80-pair measured −15% code / −10.3% cost (only solid saving in series), requires plugin hook not plain skill install
  • skills-re — skills.re — open registry of agent skills; CLI install per agent, public diffable reviews, evals per push; ~7,380 indexed skills
  • openagentskill — OpenAgentSkill — skill registry + agent-facing recommendation API (task→skill+trust+install plan); 20,842 skills, 104 agent surfaces

Concepts (77)

  • constraint-based-prompting — Constraint-Based Prompting — explicit output-format, reasoning, and behavioral constraints instead of inferred limits

  • persona-engineering — Persona Engineering — role-based prompting / persona assignment to prime knowledge and behavior subspaces; less reliable than explicit task specification

  • graphify-skill — /graphify skill: local tree-sitter AST graph of codebase+docs+media, queryable w/ file:line citations, MCP server, always-on hooks/AGENTS.md, graphifyy PyPI package

  • tool-search-tool — Claude API on-demand tool discovery: defer_loading, tool_reference expansion, regex/BM25 variants; cuts multiserver context ~85%

  • tool-context-management — four complementary strategies for tool-use context bloat: tool search, programmatic tool calling, prompt caching, context editing

  • message-batch-processing — Message Batches API: async bulk processing at 50% cost, 300k extended output, cache stacking, pause_turn continuation

  • context-compaction — checkpoint-based context management: lossy summary + retained tail, preflight trigger, instruction epoch

  • prompt-engineering-resources — catálogo de cursos, datasets, notebooks, papers, leituras, serviços e ferramentas

  • active-prompt — Active-Prompt: uncertainty-guided example selection for CoT

  • action-safety-tiering — Action Safety Tiering — risk-based confirmation policy: free local/reversible actions, confirm irreversible/shared-state ones

  • ai-agents-guide — AI agent systems: components, workflows, context engineering

  • anti-prompt-injection — Anti-Prompt Injection — defense strategies against prompt injection attacks on LLM systems

  • anti-spill-decoy — Anti-Spill Decoy — decoy injection markers to misdirect prompt extraction attempts

  • automatic-prompt-engineer — APE: LLM-generated and scored candidate instructions

  • automatic-reasoning-tool-use — ART: automatic reasoning + tool-use via frozen LLM

  • capability-as-method-calls — Capability as Method Calls — capabilities exposed as typed method calls (agent.searchGoogle()) instead of tool schemas; execution (rate limiting, retries, auth) runs outside the prompt; third answer to schema bloat after client-side deferral and gateway orchestration

  • chain-of-thought — Chain-of-Thought: step-by-step reasoning through intermediate steps

  • chain-of-thought-pattern — Chain-of-Thought Pattern — structured reasoning workflow with intermediate steps

  • clarifying-questions — Clarifying Questions — bounded (≤3) question protocol before expensive operations; prevents clarification loops

  • claude-code-environment-variables — Claude Code Environment Variables — env-var surface for auth, model selection, token budgets, caching, subagents

  • claude-code-best-practices — Claude Code Best Practices — official workflow: runnable verification, explore→plan→code→commit, aggressive context management, layered config, horizontal scaling

  • claude-code-context-window — Claude Code Context Window — startup context, compaction survival (5k/skill, 25k total caps), /compact//clear, 1M-token model gate, /context diagnostic

  • claude-code-memory — Claude Code Memory — CLAUDE.md + auto-memory systems loaded as context each session

  • claude-code-model-configuration — Claude Code Model Configuration — aliases (provider-dependent), opusplan hybrid, precedence chain, availableModels allowlist, fallback chains

  • claude-code-parallel-agents — Claude Code Parallel Agents — four parallelization approaches (subagents, agent view, agent teams, dynamic workflows), worktree isolation, token-cost multiplier

  • claude-code-sessions — Claude Code Sessions — resume/branch/export session model; /clear, /compact, /context

  • claude-code-subagents — Claude Code Subagents — markdown+frontmatter subagents in isolated context windows; five-scope precedence, model routing to Haiku, tool allow/deny

  • codebase-memory — Codebase-Memory — Tree-Sitter knowledge-graph memory over MCP replacing file reads/grep; 83% vs 92% quality at ~10× fewer tokens (paper) / 99.2% reduction claim (repo)

  • code-mode-orchestration — Code Mode Orchestration — gateway-layer pattern: MCP tool surfaces as Python stubs the model reads on demand, scripts run in a sandbox; cost bounded by what the model reads, not tool count

  • communication-style-spectrum — Communication Style Spectrum — taxonomy of output styles from solution-only to explain-before-do

  • contextsniper — ContextSniper (AntTrail) — intention-aware context gating returning recoverable evidence packets; −51.5% tokens / −36.4% cost (OpenClaw), −38.9% / −27.3% (Claude Code) with resolution unchanged; dual-source (arXiv + GitHub repo)

  • cooldown-heuristics — Cooldown Heuristics — timing and repetition guards to prevent runaway LLM loops

  • directional-stimulus-prompting — DSP: policy LM generates hints for frozen LLM

  • documented-research-process — Documented Research Process — five-phase research workflow (TDP/MIG/CAS/RG/IR) whose artifacts-as-output make process transparency structurally enforced

  • few-shot-prompting — Few-shot: in-context learning with demonstrations

  • format-interoperability — Format Interoperability — converting between XML, markdown, YAML, and JSON envelope formats

  • generate-knowledge-prompting — Generate Knowledge: LLM generates facts before answering

  • gauntlet-loop-prompting — three-part prompt structure: task + build method + quality bar; worker/critic subagent pairs looping until threshold met

  • graph-prompting — GraphPrompt: prompting framework for graph learning tasks

  • llm-models-guide — Model-specific guidance for 21+ LLMs

  • llm-research-findings — 13 research topics from agents to tokenization

  • llm-settings-and-parameters — Temperature, top-p, and inference parameters

  • markdown-header-format — Markdown Header Format — section-header-based prompt structure using and delineation

  • message-roles — Message Role Hierarchy — developer/user/assistant priority chain and the instructions parameter

  • monolithic-prompt-limits — Monolithic Prompt Limits — scaling challenges and degradation patterns in single-prompt architectures

  • multi-file-agent-architecture — Multi-File Agent Architecture — distributed prompt design across multiple specialized files

  • multimodal-cot — Multimodal CoT: two-stage vision+text reasoning

  • paired-thinking-response — Paired Thinking Response — dual-output pattern: internal reasoning + user-facing answer

  • pal-program-aided-language-models — PAL: programs as intermediate reasoning (Python)

  • plan-before-implement-gate — Plan-Before-Implement Gate — gate pattern requiring explicit planning before action execution

  • preliminary-context-gathering — Preliminary Context Gathering — pre-response phase for collecting project context and requirements

  • prompt-chaining — Prompt Chaining: sequential sub-prompts for complex tasks

  • prompt-caching — Prompt Caching — API cache_control mechanism: 0.1× reads, 1.25–2× writes, model-specific prefix minimums

  • prompt-engineering-applications — Real-world LLM applications: data gen, code gen, function calling, RAG

  • prompt-engineering-guide — Prompt Engineering Guide (promptingguide.ai) — comprehensive educational resource

  • prompt-engineering-guides — Practical tutorials: optimizing, deep research, reasoning

  • prompt-engineering-risks — Adversarial prompting, factuality, and bias risks

  • prompt-engineering-techniques — 18 prompting techniques from zero-shot to GraphPrompt

  • prompt-hub-catalog — 25+ categorized prompt examples

  • rag-prompting — RAG as a prompting technique: retrieval-augmented generation at inference time

  • react-prompting — ReAct: interleaved reasoning traces + task actions

  • reasoning-effort — Reasoning Effort and Mode — API-level reasoning control (effort/mode/context) on OpenAI reasoning models

  • reflexion — Reflexion: self-correction via verbal reinforcement feedback

  • self-consistency — Self-Consistency: sampling multiple CoT paths, majority voting

  • skill-reducer — SkillReducer — two-stage skill optimization: 48% description / 39% body compression with +2.8% functional quality (55,315-skill study)

  • skills-directory-gate — Skills Directory Gate — capability enumeration and routing using a centralized skills manifest

  • skill-authoring-best-practices — Skill Authoring Best Practices (Anthropic) — concise SKILL.md, 3rd-person descriptions, degrees of freedom high/medium/low, progressive disclosure

  • structured-outputs — Structured Outputs — JSON-Schema-guaranteed response format via text.format json_schema

  • task-state-machine — Task State Machine — finite-state-machine pattern for structured multi-step workflows

  • template-variable-prompts — Template Variable Prompts — prompt templates with {{variable}} substitution for dynamic contexts

  • token-usage-reduction — Token Usage Reduction — context management, caching, model selection, rate-limit tiers, and compression-tool caveats for cutting token spend

  • token-waste-sources — Token Waste Sources — diagnostic taxonomy of where agents waste context (terminal logs, full files, grep loops, old state, verbosity, over-built code) + measurement checklist

  • token-reduction-is-not-cost-reduction — Token Reduction Is Not Cost Reduction (Paper) — 2,908-run measured study: −38% tool-output tokens → +6.8% billed cost; cache ~87% of cost; success-adjusted billed cost as the metric

  • tree-of-thoughts — Tree of Thoughts: BFS/DFS search over reasoning states

  • two-tier-search-architecture — Separation of planning/search from answer generation

  • tool-priority-override — Tool Priority Override — system-prompt pattern where one tool takes absolute priority over all others

  • xml-envelope-format — XML Envelope Format — system prompt organized within tags for hierarchical structure

  • zero-shot-prompting — Zero-shot: LLMs performing tasks without examples

Comparisons (5)

Queries (6)

Sources

See root AGENTS.md and prompt wiki AGENTS.md for operating rules.

Visão de gráfico

Backlinks