mk:llms

llms.txt Generator

Generate llms.txt files — LLM-friendly markdown indexes of project documentation.

Path convention: Commands below assume cwd is $CLAUDE_PROJECT_DIR (project root). Prefix paths with "$CLAUDE_PROJECT_DIR/" when invoking from subdirectories.

When to Use

• Project needs LLM-friendly documentation index
• User asks for "llms.txt", "LLM documentation", "AI-friendly docs"
• Publishing docs site and want AI discoverability
• After major documentation updates (Phase 6 Reflect)

Explicit: /mk:llms [path] [--full]

Workflow Integration

Operates in Phase 6 (Reflect) or on-demand. Output supports the documenter agent.

Arguments

| Flag | Effect | | --------------- | ------------------------------------------------- | | (no args) | Scan ./docs directory | | path | Scan specific directory | | --full | Also generate llms-full.txt with inline content | | --output path | Custom output location (default: project root) | | --url base | Base URL prefix for links |

Scripts

Script-first approach — the Python script handles all deterministic work (scanning, extracting, categorizing, generating). Claude only reviews and improves the output.

# Generate llms.txt (script does the heavy lifting)
.claude/skills/.venv/bin/python3 .claude/skills/llms/scripts/generate-llms-txt.py \
  --source ./docs [--output .] [--base-url https://example.com/docs] [--full]

# Preview metadata first (Claude reviews before generating)
.claude/skills/.venv/bin/python3 .claude/skills/llms/scripts/generate-llms-txt.py \
  --source ./docs --json

Process

1. Run script — execute generate-llms-txt.py --source <path> --json to preview metadata
2. Review metadata — check project name, description, file categorization
3. Generate — run script again without --json to produce llms.txt
4. Review output — read generated llms.txt, improve descriptions if needed
5. Validate — check: H1 present, blockquote summary, valid link format, Optional section last
6. If --full: review llms-full.txt for completeness

Output Format

# {Project Name}

> {Brief project description with essential context.}

## {Section Name}

- [{Doc Title}]({url}): {Brief description}
- [{Another Doc}]({url}): {What this covers}

## Optional

- [{Less Important Doc}]({url}): {Supplementary info}

References

| Reference | When to load | Content | | ----------------------------------------------------- | ------------------- | ------------------------------------ | | [llms-txt-spec.md](./references/llms-txt-spec.md) | Step 4 — generating | Full llmstxt.org specification rules |

Failure Handling

| Failure | Recovery | | ------------------------ | ----------------------------------------- | | No .md files found | Report "No documentation found in {path}" | | No H1 in doc file | Use filename as title, warn user | | Output path not writable | Suggest alternative path |

Constraints

• Never fabricate documentation content — only index what exists
• Follow llmstxt.org spec strictly
• Keep descriptions concise — one sentence per entry

Gotchas

• Python venv required: run .claude/scripts/bin/setup-workflow once from the project root before invoking this skill. Re-run after upgrading the toolkit.