Immutable. This exact content is served forever at /api/v1/blob/c87ecc2268395dce.
---
name: make-docs
description: >-
Use when the user wants architecture docs or behavioral specs, asks to explore a codebase for documentation, sync docs after changes, or record an ADR. Branches: explore (survey, propose, generate), update (since last docs stamp → sync), adr (decision record).
metadata:
version: 0.1.0
author: "Diego Oliveira"
tags:
- docs
- architecture
- specs
- adr
---
# Make docs
Produces a documentation suite in `docs/`: **architecture** (how it's structured) and **specs** (what it does). Specs name only **observable** outcomes — an agent can implement or verify against them.
## Commands
- **`explore`** — survey codebase, propose structure, generate full suite
- **`update`** — sync docs with commits since the last docs stamp
- **`adr`** — record one architecture decision
## Layout
```text
docs/
README.md # includes `> Updated on <ISO-date>`
architecture/
architecture.md
domains/
glossary.md
api.md # only if the project exposes an API
constraints.md
decisions/ # ADR-NNNN-<slug>.md
patterns.md
specs/
<domain>.md # one per behavioral domain
```
By repo size: **small** — README + architecture.md + decisions/; **medium/large** — full set; **monorepo** — root docs/ + one per service (nearest wins). Cut any file or section the survey did not **earn**. Fill every earned section from code evidence — no template stubs left behind.
## Templates
Before writing a file, load its template from `references/`:
| Template | File |
| ------------------------------------- | ------------------------------------------ |
| `references/template-readme.md` | `docs/README.md` |
| `references/template-architecture.md` | `docs/architecture/architecture.md` |
| `references/template-glossary.md` | `docs/architecture/domains/glossary.md` |
| `references/template-api.md` | `docs/architecture/domains/api.md` |
| `references/template-constraints.md` | `docs/architecture/domains/constraints.md` |
| `references/template-adr.md` | `docs/architecture/decisions/ADR-*.md` |
| `references/template-patterns.md` | `docs/architecture/patterns.md` |
| `references/template-spec.md` | `docs/specs/<domain>.md` |
## Specs
Specs describe **observable** behavior — outcomes a user or external system can see. Name domains after behavior (`auth`, `orders`), not folder layout.
- One requirement = one **SHALL** (split compound requirements).
- Keywords: MUST / SHALL / SHOULD / MAY (RFC 2119).
- Each requirement has at least one Scenario (Given/When/Then) that **exercises** it — a concrete case, not a rewording.
- Cover the case that would hurt most to see broken; name it explicitly.
- Pure refactor with no behavior change → no spec update (record that).
## explore
1. **Survey** — stack, runtime, entry points, data stores, API protocol (if any), auth (if any), deployment, patterns, quality targets. Identify behavioral domains by feature. Read existing docs first.
**Done when:** stack, runtime, entry points, and deploy named from code; auth named only if present; every behavioral domain listed by feature name; existing docs read; out-of-scope items flagged or skipped.
2. **Propose** — list files to create; present and ask for approval. Unattended (Agent mode): assume the most reasonable list, proceed, and record the assumption.
**Done when:** user confirmed or declined, or assumption recorded.
3. **Generate** — for each file, load its template and write. Write README.md last (indexes every other file; set `> Updated on` to today's ISO date). Cut what the survey did not **earn**. Fill every earned section from evidence.
**Done when:** every confirmed file written from its template; unearned sections omitted; no placeholder stubs in earned sections; stamp set.
4. **Verify** — claims confirmed against code; every ADR names a _why_; every requirement has a scenario; quality targets are numbers; diagrams match the code they summarize; README indexes every file; earned sections are filled from evidence.
**Done when:** every check above passes.
## update
1. **Scope** — read `docs/README.md` for `> Updated on <ISO-date>`. Count commits since that date (`git rev-list --count HEAD --since=<date>`).
- Missing stamp → ask the user for the range (or offer a full re-sync).
- Count ≤ 10 → cover all commits since the stamp.
- Count > 10 → inform the count and ask: cover all, last N commits, or a specific base (SHA / date). Wait for the answer before diffing.
**Done when:** base revision chosen (all since stamp, last N, or user-specified base).
2. **Diff** — `git diff <base>..HEAD` plus unstaged/uncommitted changes → classify: structural → architecture, behavioral → specs, decision → ADR.
**Done when:** every changed path classified.
3. **Update** affected files. Load the template before writing. Pure refactor → record "no spec change" explicitly. Set `> Updated on` in `docs/README.md` to today's ISO date.
**Done when:** every classified change reflected or explicitly waived; stamp refreshed.
4. **Verify** — every behavioral change has spec update or "no spec change"; structural changes in architecture.md; diagrams match code; cross-references resolve; stamp present and current.
**Done when:** every check above passes.
## adr
Load `references/template-adr.md`. Number `ADR-NNNN-<slug>.md`. Supersede with two-way links. If structural boundaries change, update architecture.md.
**Done when:** names a _why_; gains + costs present; alternatives have reasons; architecture.md updated if boundaries moved.
## Out of scope
Data model, observability, CI/CD. If detected during survey, ask before generating; otherwise skip.