Docs From Prs

docs-from-prs

Inputs

Optional args (free-form, parse from the user's message):

• since: a date (2026-04-01), a PR number (#24), or a count (last 15). Default: PRs merged after the most recent commit that touched the primary doc (README.md by default), capped at 20.
• remotes: origin or a named remote, or all if the project mirrors to multiple GitHub orgs. Default: the current repo's origin. Multi-remote projects should follow their CLAUDE.md / AGENTS.md PR workflow.
• dry-run: report the gap analysis only, no edits. Use when the user says "what's missing" or "audit".

Step 1 — Survey

Resolve the current repo first:

REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)

Then run these in parallel:

# Recent merged PRs
gh pr list --repo "$REPO" --state merged --limit 20 \
  --json number,title,mergedAt,body,files

# Last commit that meaningfully touched the primary docs
git log -1 --format='%H %ad' --date=short -- README.md

# Recent commits for context
git log --oneline --since="<derived date>" | head -40

If gh pr list 404s, the wrong gh account is active — switch with gh auth switch --user <handle>. The project's CLAUDE.md or AGENTS.md should name the handle for each remote.

Step 2 — Classify each PR

For each PR, decide one of:

• user-facing — flag/CLI/alias → CLI reference + alias/command table + changelog + user guide
• user-facing — config/env → config example block + user guide config section
• user-facing — behaviour/fix → "Recent Changes" + changelog detail entry + (optional) "How it works" subsection
• internal chore → skip (CI tweaks, refactors, dependency bumps, agent config moves)
• already covered → skip (verify with grep against current docs before skipping)

Read the actual source — don't trust PR titles alone. Use git show <sha> --stat and inspect:

• New flags: command-entry scripts / CLI parser modules
• New aliases or shortcuts: any aliases.sh, shell wrappers, package.json scripts, Justfile, Makefile
• New config: .<tool>rc files, env-var loader modules, schema definitions
• New subcommands or modes: dispatcher / router / main entry point

Step 3 — Drift hot spot audit (mandatory, every run)

Most projects have one or two surfaces that drift faster than the rest — alias tables, CLI flag matrices, env-var lists, supported-version tables. Identify these in the project's docs and audit them every run, even if no PR explicitly touches them.

A reasonable audit pass:

1. Enumerate the source of truth (grep -nE '^(alias |[A-Za-z_]+\(\) \{$)' aliases.sh, --help output, schema file, package.json scripts, etc.).
2. Cross-check each entry against the doc tables. New entries → add a row. Renamed/removed → fix or delete.
3. Spot-check the variants that historically drift first. If the project's CLAUDE.md / AGENTS.md names the hot spots, treat that list as authoritative; otherwise ask the user which surfaces tend to drift.

Drift accumulates from minor commits. Run this audit even when the survey returns zero PRs.

Step 4 — Section placement

Place each gap thoughtfully by change type — don't dump everything into "Recent Changes":

Identify the project's actual doc layout before placing. Common layouts:

• Single-file (README.md only) — collapse columns 2 and 3 into the README.
• README + landing page (docs/index.html, index.html, GitHub Pages site) — landing pages often carry a JS-driven changelog or detail panel; check for a CHANGE_DETAILS / changelog map and shift numeric keys when inserting items at the top.
• README + user guide (docs/, wiki/, book/) — multiple pages; put narrative explanation in the guide, terse reference in the README.
• README + site + user guide — apply the full matrix above.

Out-of-scope directories vary — check the project's CLAUDE.md / AGENTS.md for "never edit" lists (commonly docs/archive/, docs/generated/, docs/code-review/, vendored docs).

Step 5 — Edits

• Additive by default: leave working prose as-is.
• Match surrounding voice and emoji policy (some pages use emojis, others don't — stay consistent within each file).
• Read every file before editing; for large reference files (>30KB) use grep -n to find anchors instead of full reads.
• Verify HTML stays well-formed after editing (tag balance + any JS-driven detail map).
• Leave generated artefacts untouched (e.g. graphify-out/, autogenerated API docs, build outputs).

Step 6 — Copy-edit pass (mandatory, after content edits, before commit)

After all content additions are in place, do a grammar/style/alignment pass over only the files you touched this run (don't churn unrelated prose).

Grammar and prose

• Subject-verb agreement, article use (a/an/the), pluralization, tense consistency.
• No double spaces, no trailing whitespace, no smart quotes mixed with straight quotes.
• Oxford comma when a list has three or more items, matching the rest of the doc.
• Active voice where the surrounding doc uses active voice; don't flip an existing passive section.
• Hyphenation: command-line (adj), command line (noun); multi-repo, cross-repo, worktree (one word). Treat literal filenames as inviolable tokens — never reflow them across a line.
• Code identifiers, flags, env vars, file paths in backticks.

Title casing

Match the dominant casing of the file you are editing:

• Sentence case (## How it works) — most modern READMEs and Markdown user guides.
• Title case (## How It Works) — many landing-page <h1>/<h2> headings, some legacy READMEs.

Do NOT normalize across the whole doc — match the immediate neighbourhood. When in doubt, run grep -E '^#{1,4} ' <file> | head -30 and copy the dominant pattern.

Other casing rules:

• Table column headers: sentence case unless the column is a single proper-noun token.
• Inline emphasis labels (**Why:**, **How to apply:**): keep the exact casing already established.
• Proper nouns always capitalized: project names, language names, GitHub, CLI, README, etc.

Alignment and layout

• Markdown tables: re-align pipes after edits. Every column's pipe must line up; pad with spaces. Header separator row matches column widths.

  | Column A | Column B     | Column C |
  |----------|--------------|----------|
  | value    | longer value | x        |
  

• Lists: consistent bullet character (- not * if the file uses -); two-space indent for nested lists; blank line before and after a list block.
• Code fences: language tag on every fence (```bash, ```html, ```text for plain). No bare ``` if the rest of the file tags them.
• Blank lines: exactly one blank line between sections; no triple-blank gaps; file ends with a single trailing newline.
• Line length: don't hard-wrap prose paragraphs unless the file already hard-wraps. Don't reflow existing paragraphs you didn't otherwise edit.

HTML-specific (when editing landing pages)

• Tag balance: every <li>, <div>, <section> you opened is closed. Verify with grep -c '<li>' <file> vs grep -c '</li>' <file>.
• Indentation: match the surrounding 2-space or 4-space style — don't introduce a new style.
• JS-driven changelog maps (e.g. CHANGE_DETAILS['latest']): keys are sequential integers. After inserting items at top, every existing key must shift by the count of inserted items. Verify by counting <li> items vs object keys.

Verification commands

Run these after the edit pass:

# Trailing whitespace / mid-line double spaces
grep -nE ' +$' <files-you-touched>
grep -nE '  +[^ ]' <files-you-touched>

# Heading casing audit (skim output, don't blindly fix)
grep -nE '^#{1,4} ' <files-you-touched>

# HTML tag balance for changelog
grep -c '<li>' <html-file>
grep -c '</li>' <html-file>

If any check surfaces drift introduced by this run, fix it. If the drift predates this run, note it in the report but don't fix it (out of scope — separate cleanup PR).

Step 7 — Commit, branch, PRs

Branch name pattern: docs/update-recent-prs (or docs/update-<topic> if the survey covers a single theme).

git checkout -b docs/update-recent-prs
git add <doc-files>
git commit -m "docs: cover recent PRs (#X-#Y)"

PR creation:

• Default: one PR to the project's primary base branch (main or dev, per project convention).
• Multi-remote projects (mirrored to several GitHub orgs): follow the project's CLAUDE.md / AGENTS.md PR workflow. Switch gh accounts between remotes with gh auth switch --user <handle>.
• Use absolute paths with git -C <worktree-path> if running from a worktree.

Step 8 — Report

Return a structured summary:

• Branch name, commit SHA
• Files changed (bulleted)
• One line per PR: gap-filled / already-covered / skipped (with reason)
• Drift hot spot audit result: rows added / corrected / clean
• PR URLs if created
• Copy-edit pass: items fixed (grammar, casing, alignment) vs items flagged as pre-existing drift
• Any factual mismatches discovered between docs and source (these are bugs, flag them even if out of scope)

Constraints

• Never push without confirming. Default is push branch + open PR(s); destructive force-push or rewriting history requires explicit user ask.
• Never edit out-of-scope directories. Check CLAUDE.md / AGENTS.md for the project's "never edit" list — typical entries: docs/archive/, docs/generated/, docs/code-review/, vendored docs.
• Never invent flags or aliases — read the code.
• If a PR's claimed feature can't be found in the source, surface it in the report rather than documenting a ghost feature.