doxcavate

Produces durable, structured documentation in codebases where written docs are sparse — readable by humans and agents alike.

Install

npx skills add xornivore/skills@doxcavate --agent claude-code -y

When to use

Invoke when the user:

• Asks to inventory or document a codebase ("doxcavate this repo", "what should be documented?", "doxcavate the X") — route to survey.
• Names a specific doc to write or refresh ("doxcavate runbook-deploy", "write a how-it-works for the ingestion pipeline") — route to draft.
• Asks to update existing docs after code changes — route to draft.
• Phrasing is ambiguous — ask once, then route.

Do not invoke for:

• Auto-generated API references (OpenAPI → Markdown, godoc, etc.) — tool-shaped, not doc-shaped.
• ADRs / decision records — they need commit-time discipline this skill can't manufacture after the fact.

Hard rules

1. Never silently write outside the host repo. partial and shadow storage modes are opt-in and require user acknowledgment on first write. Audit: any write to ~/.local/share/doxcavate/ must be preceded by a logged mode-confirmation prompt.
2. Never silently add meta-files (.doxcavate.yml, docs/) to a repo that has none. Ask first. Audit: git diff after a first invocation must not introduce these without a logged user acknowledgment.
3. In partial mode, leaves committed to the repo must not link to shadow-located meta docs. Shadow paths are per-machine and per-user; rendering them into a committed leaf would break for everyone else. Leaf-to-leaf links inside the repo are fine. Audit: grep -rE '~/\.local/|/Users/[^/]+/\.local/' over the repo's committed leaves returns no matches.
4. Code wins on facts. When sources disagree, the code is the ground truth. See investigation-and-sources.
5. Factcheck is non-skippable. No flag, prompt hint, or config override skips it. Audit: every produced doc carries a ## Verification block populated by the factcheck pass.
6. Persona output is capped at 7 ranked items. No exceptions. If the reviewer would have to drop substantive items to fit, it emits E_PERSONA_OVERFLOW and halts. See persona-review.
7. No git blame author attribution in any doc this skill writes. Audit: grep -rE '@[a-zA-Z0-9_-]+|<[^>]+@[^>]+>' over produced docs returns no author tokens.

Reference index

Read each reference only at the action step that calls for it. Each reference ends with ## Action checklist — load and jump there.

• layout-and-discovery — config sniffing, the three storage modes (integrated, partial, shadow), Mode-by-signal table, repo keying. Read at discovery (always step 1).
• invocation-modes — survey vs draft, routing details, index path by storage mode. Read once mode is decided.
• doc-kinds — taxonomy, canonical front-matter schema, required structural anchors per kind, sizing targets. Read when drafting or reviewing structural conformance.
• investigation-and-sources — adaptive investigation, the three production sources (code, existing docs, commits), reconciliation order. Read at the start of any draft.
• factcheck — claim extraction, verification, ## Verification block as byproduct, E_TOO_MANY_CONTRADICTIONS halt. Read at the factcheck step.
• persona-review — kind-keyed persona review, output contract, quality bar, E_PERSONA_OVERFLOW halt. Read at the persona-review step.
• review-loop — orchestration of factcheck and persona, single-round default, opt-in thorough mode. Read once the prior two are loaded.

Persona briefs are prompt templates, loaded from assets/persona-prompts/<kind>.md at the persona-dispatch step.

Top-level action checklist (when invoked)

1. Run discovery: load layout-and-discovery, match the Mode-by-signal table, resolve config / docs root / storage mode.
2. Route survey vs draft per the When to use rules above.
3. Survey: load invocation-modes, inventory the repo (see investigation-and-sources), propose the doc plan, write index.md to the resolved index path, stop.
4. Draft:
   1. Load index.md from the resolved path.
   2. Run the production-sources flow per investigation-and-sources.
   3. Copy assets/<kind>-template.md for the target doc kind (see doc-kinds).
   4. Fill placeholders using the reconciled sources.
   5. Run factcheck per factcheck (never skippable).
   6. Run persona review per persona-review, quoting assets/persona-prompts/<kind>.md in the dispatch.
   7. Apply all returned items, preferring patch over textual instruction.
   8. Run focused factcheck on changed regions only.
   9. Done — or, in thorough mode (per review-loop), repeat steps 5–8 up to 3 total rounds.