docforge-llms — sub-skill of docforge

Canonical Markdown → AI-native projections (llms.txt, AGENTS.md, CLAUDE.md, MCP, .cursor/rules, .github/copilot-instructions, per-page .md).

One canonical. Many projections. Audience declared, or refused.

Layout

docforge-llms/
├── SKILL.md                  # Body. ≤500 lines. Lazy-loads references.
├── README.md                 # This file.
├── references/               # Progressive-disclosure deepdives (load on demand).
│   ├── llms-txt-spec.md
│   ├── claude-md-anthropic.md
│   ├── agents-md-spec.md
│   ├── mcp-spec-essentials.md
│   ├── cache-economics.md
│   ├── chunkability-rules.md
│   └── dual-output-pattern.md
├── templates/                # Output templates with deterministic placeholders.
│   ├── llms-txt.template
│   ├── llms-full-segmented.template
│   ├── claude-md.template
│   ├── agents-md.template
│   ├── cursor-rules.template.mdc
│   ├── copilot-instructions.template.md
│   └── mcp-server.template.py
├── scripts/                  # Capability implementations.
│   ├── project.py            # canonical → per-page .md + llms.txt
│   ├── forge_agents_md.py
│   ├── forge_claude_md.py
│   ├── mcp_scaffold.py
│   ├── chunkability_lint.py
│   ├── cache_optimize.py
│   ├── diagram_alt_referencer.py
│   └── lib/
│       ├── frontmatter.py
│       ├── tone_transformer.py
│       └── token_counter.py
├── audiences/                # Per-audience profiles (knobs for tone transformer).
│   ├── llm-runtime.profile.yaml
│   ├── llm-rag.profile.yaml
│   ├── human-novice.profile.yaml
│   ├── human-expert.profile.yaml
│   └── ci.profile.yaml
└── tests/                    # Smoke tests.

Capabilities (v0.1)

Invariants

I1 audience-declared · I2 canonical=MD-in-git · I3 CLAUDE.md ≤200 lines · I4 AGENTS.md is canonical · I5 llms.txt = Howard spec · I6 MCP typed · I7 cache-aware order · I8 provenance tag · I9 no human edits a projection.

See SKILL.md §1 for the full table.

Sources (verbatim references)

• llms.txt spec — Howard / Answer.AI 2024 → references/llms-txt-spec.md
• CLAUDE.md — Anthropic verbatim → references/claude-md-anthropic.md
• AGENTS.md — AAIF / Linux Foundation Dec 2025 → references/agents-md-spec.md
• MCP — modelcontextprotocol.io 2025-11-25 → references/mcp-spec-essentials.md
• Cache economics — Anthropic prompt-caching docs → references/cache-economics.md

Refuses

• No audience declared → refuse.
• CLAUDE.md target >200 lines → refuse.
• forge-claude-md without AGENTS.md → refuse.
• Edit a projection by hand → refuse. Edit canonical, regenerate.

<!-- generated:llm reviewed:human@HEAD -->

---

_{Built with MyClaude Studio Engine}