Documenting Projects

Announce: "Using documenting-projects skill for documentation generation."

<BEHAVIORAL_MODE> ORCHESTRATOR: Dispatch subagents via Task tool for ALL documentation work. Each subagent invokes its phase command via the Skill tool (e.g., Skill("docs-audit")). Never write documentation directly. Never invoke phase commands directly from the orchestrator. Context should contain only Task dispatches, user communication, and phase transitions. </BEHAVIORAL_MODE>

Invariant Principles

1. Diataxis First: Every document maps to exactly one Diataxis type. Mode mixing is a defect.
2. Tone is Non-Negotiable: Each section follows its assigned tone profile. Deviations trigger re-generation.
3. Anti-AI-Tone Rules Apply to ALL Content: No banned phrases, no hedging, no marketing language. Enforced during generation, not as an afterthought.
4. Quality Gates Must Pass Before Docs Ship: All 8 review criteria evaluated, all must pass or user explicitly accepts known issues.
5. Output Goes to Project Repo: Documentation files are project deliverables, not spellbook artifacts. Write directly to the project's docs directory.

Phase Overview

Phase Transition Diagram

/docs-audit --> audit-result.json --> /docs-plan --> plan.json --> /docs-write --> written-manifest.json --> /docs-review
                                                                       ^                                        |
                                                                       +---- sections-filter.json <-------------+
                                                                             (iteration loop, max 2)

Phase 1: Audit

Execute: Dispatch subagent via Task tool. Subagent invokes Skill("docs-audit").

Outputs:

• doc-state/audit-result.json (DocsAuditResult)
• User-facing summary of project analysis

Self-Check:

• audit-result.json exists and contains all required fields
• User has chosen existing docs mode (audit+improve / start fresh / additive)

Phase 2: Plan

Execute: Dispatch subagent via Task tool. Subagent invokes Skill("docs-plan").

Outputs:

• doc-state/plan.json (DocsPlan)
• Build config file (e.g., mkdocs.yml)

Self-Check:

• plan.json exists with TOC, build config, and tone assignments
• User has explicitly approved the TOC and build config

Phase 3: Write

Execute: Dispatch subagent via Task tool. Subagent invokes Skill("docs-write").

When sections-filter.json exists (iteration from Phase 4), pass it to the subagent. The writing phase regenerates only the specified sections and merges into the existing manifest.

Outputs:

• Documentation files written to project repo
• doc-state/written-manifest.json (DocsWritten)

Self-Check:

• written-manifest.json exists
• All MVP sections from plan.json are accounted for (written or explicitly skipped)

Phase 4: Review

Execute: Dispatch subagent via Task tool. Subagent invokes Skill("docs-review").

Outputs:

• doc-state/review-result.json (DocsReview)

Self-Check:

• review-result.json exists with all 8 criteria evaluated
• If overall_pass == true: announce completion, offer hosting config
• If overall_pass == false and iteration_count < 2: write sections-filter.json, re-dispatch Phase 3 (/docs-write), then re-dispatch Phase 4 (/docs-review)
• If overall_pass == false and iteration_count >= 2: escalate to user

Standalone: /write-readme

When user says "write a README", "create README", "update README", or invokes /write-readme:

1. Skip this orchestrator entirely
2. Invoke /write-readme command directly
3. No doc-state involvement, no multi-phase pipeline

The /write-readme command performs its own lightweight project analysis and generates a README with progressive disclosure structure. Use it for single-file README work.

Session State

All phase outputs are stored as JSON at ~/.local/spellbook/docs/{project-encoded}/doc-state/:

Iteration state: When /docs-review identifies failing sections, it writes sections-filter.json (a string[] of output paths to regenerate). The orchestrator then re-dispatches /docs-write, which reads sections-filter.json to regenerate only those sections.

Re-run Policy

Each phase checks whether its output file already exists at the doc-state path. If found, present to the user: "Previous [phase] results found at [path]. Reuse existing results or regenerate?"

• Reuse: Skip the phase, proceed with existing data.
• Regenerate: Overwrite the file and re-run the phase.

Quality Gate Thresholds

The 8 quality criteria evaluated in Phase 4:

Subagent Result Schemas

Each Task dispatch MUST specify a strict JSON result schema.

Explore agent (audit phase):

{
  "language": "string",
  "framework": "string",
  "build_tool_recommendation": "string",
  "existing_docs": [{"path": "string", "type": "string", "quality": "string", "staleness": "string"}],
  "api_surface": {"modules": ["string"], "coverage": "number"},
  "gaps": ["string"]
}

Writing subagent:

{
  "files_written": [{"path": "string", "diataxis_type": "string", "tone_profile": "string", "word_count": "number", "last_verified_date": "string"}]
}

Review subagent:

{
  "results": [{"criterion": "string", "passed": "boolean", "details": "string", "file": "string"}],
  "overall_pass": "boolean"
}

Fact-checking subagent:

{
  "claims": [{"claim": "string", "verdict": "verified | unverified | false", "evidence": "string"}]
}

Execution Protocol

Phase 1: Dispatch subagent -> Skill("docs-audit") -> verify audit-result.json exists
Phase 2: Dispatch subagent -> Skill("docs-plan") -> verify plan.json exists
Phase 3: Dispatch subagent -> Skill("docs-write") -> verify written-manifest.json exists
Phase 4: Dispatch subagent -> Skill("docs-review") -> check overall_pass
  If !overall_pass && iteration < 2:
    docs-review writes sections-filter.json from failing_sections
    Re-dispatch Phase 3 (docs-write reads sections-filter.json)
    Re-dispatch Phase 4 (docs-review)
  If !overall_pass && iteration >= 2:
    Escalate to user with specific failures
    User decides: accept with known issues OR manual fix
  If overall_pass:
    Announce completion
    Offer hosting config (GitHub Pages, ReadTheDocs, Netlify)

MVP Scope

Tier 1 (MVP): README + Tutorials + Reference + MkDocs build tool.

Dispatch Model

The orchestrator dispatches subagents via the Task tool. Each subagent prompt provides CONTEXT (project path, state file paths, user preferences) but never duplicates command instructions. The subagent loads the command via Skill("docs-audit") etc. and follows the command's own logic.

Subagent prompts should include:

• Project root path
• Doc-state directory path
• Any user preferences or overrides from earlier phases
• Reference to prior phase output (file path, not inline content)

Self-Check

Before declaring the documentation pipeline complete:

• Phase 1 executed: audit-result.json exists
• Phase 2 executed: plan.json exists and user approved TOC
• Phase 3 executed: written-manifest.json exists, all MVP sections accounted for
• Phase 4 executed: review-result.json exists, overall_pass == true (or user accepted)
• All documentation files written to project repo (not spellbook artifact directories)
• No iteration loops remain open

Citations

Sources that informed this skill's design:

• Procida, D. "Diataxis: A systematic approach to technical documentation." https://diataxis.fr/
• Google Developer Documentation Style Guide
• Microsoft Writing Style Guide
• Stripe API Documentation (architectural patterns for reference docs)
• React Documentation (pedagogical patterns for tutorials)
• Tailwind CSS Documentation (visual/justification patterns for how-to guides)
• Markdoc (extensibility framework reference)
• The Good Docs Project (template patterns)