Generate Docs

generate-docs

You are the orchestrator for a documentation-generation pipeline. Coordinate five phases — Bootstrap → Explore → Generate → Verify → Context Update — in an iterative loop until every document meets grade A on every rubric dimension, or the retry budget is exhausted.

Detailed instructions for each phase live in references/. Read the relevant reference file when you reach that phase (progressive disclosure — don't load them all up front).

Configuration

• Max retries: the first argument, if given (e.g. generate-docs 5). If blank or non-numeric, default to 3.
• Output directory: docs/ at the project root.
• Passing grade: A on every document and every rubric dimension.
• Completion: when ALL docs score A across ALL dimensions, or max retries is reached.

How to run the phases (any harness)

This skill is harness-agnostic. Each phase is a batch of independent work that parallelizes cleanly:

• If your harness supports subagents / parallel tasks (e.g. a Task tool, an agent-spawn primitive, or a workflow engine): dispatch the phase's work as parallel subagents — 8 explorers in Phase 1, up to 15 writers in Phase 2, 1 verifier in Phase 3. Give each subagent the matching references/ file as its instructions, plus its specific brief/doc. Prefer read-only tools for explorers and verifiers.
• If your harness has no subagents: perform each phase's work yourself, sequentially, following the same references/ files. The output is identical; only the wall-clock time differs.

Whichever mode you use, the phase boundaries, working files, and grading are the same.

Iteration loop

Phase 0: Bootstrap (run once)

iteration = 0
max_retries = first argument or 3

WHILE iteration < max_retries:
    iteration += 1

    IF iteration == 1:
        Run Phase 1 (Explore)
        Run Phase 2 (Generate ALL docs)
    ELSE:
        Run Phase 2 ONLY for docs that scored below A,
        using the verifier's feedback (revision mode). Preserve passing docs.

    Run Phase 3 (Verify)

    IF verdict == PASS:
        Run Phase 4 (Context Update); BREAK → success
    IF iteration == max_retries:
        Run Phase 4 (Context Update, even on partial pass); BREAK → partial success

Print the summary table.

Final output

When the loop ends, print a summary to the console:

📚 Documentation Generation Report
Iteration:  {n} / {max_retries}
Verdict:    {PASS ✅ | FAIL ❌}
CLAUDE.md:  {Created ✨ | Updated 🔄 | Unchanged ─}
AGENTS.md:  {Created ✨ | Updated 🔄 | Unchanged ─}

Document                     CMP ACC CLR DIA COV  Overall
README.md                    ...
architecture.md              ...
tech-stack.md                ...
api-reference.md             ...
database-schema.md           ...
frontend-backend.md          ...
services.md                  ...
workflows-and-triggers.md    ...
devops.md                    ...
glossary.md                  ...
error-catalog.md             ...
testing-strategy.md          ...
environment-setup.md         ...
coding-conventions.md        ...
security.md                  ...

Docs generated:  docs/ (15 files)
Context updated: CLAUDE.md, AGENTS.md
SpecKit-ready:   docs/ structure compatible with /speckit.analyze

On FAIL, also list the specific failing dimensions and the verifier's feedback for each doc.

Critical rules

1. Never invent information. Document only what actually exists in the codebase.
2. Always include Mermaid diagrams where the doc table marks them "Required".
3. Verify against code, not just the exploration report. The verifier reads actual source files.
4. On retries, only regenerate failing docs — never the whole set.
5. Preserve passing docs across iterations — do not overwrite docs that already scored A.
6. The exploration and verification reports are working files (docs/_*.md) — don't delete them.
7. Be honest about gaps. Flag uncertainties with ⚠️ rather than guessing.
8. CLAUDE.md stays under ~200 lines total; the auto-generated section under 80. Use progressive disclosure — point to docs, don't duplicate their content.
9. Never delete human-written CLAUDE.md content. Only manage the marked auto-generated section.
10. AGENTS.md is tool-agnostic — no Claude-specific syntax; keep it under 120 lines.
11. Structure for AI agents. Explicit headings, tables, and "Read when" triggers so SpecKit, Cursor, Copilot, and other tools can extract context efficiently.