---

name: docs-review description: Review docs and blog post quality before committing (style, accuracy, Pulumi best practices). Use when you've made content changes locally and want a quality pass on open files, the current branch, or a specific PR — outputs to the conversation, never posts to GitHub. user-invocable: true

Docs Review (interactive)

Output goes into the conversation. This skill never posts to GitHub.

This is the pre-push pass. Once a PR is open and CI has posted its pinned review, /address-review is the skill that works those findings to zero with the author.

Usage

/docs-review [PR_NUMBER]

If PR_NUMBER is provided, review the PR via gh pr view / gh pr diff. If omitted, auto-detect scope from the current IDE/terminal context.

Scope detection (when no PR number is provided)

Walk these steps in order; stop at the first that yields a scope.

1. Open files in the IDE. Check the conversation context for system reminders that list open files. If any are present, treat those files as the scope.

2. Uncommitted changes. Check the gitStatus block (or run git status) for modified (M) and untracked (??) files. Use git diff and read the affected files.

3. Branch changes vs. master:

   git diff $(git merge-base --fork-point master HEAD)...HEAD
   

   If --fork-point fails (no reflog), fall back to git diff $(git merge-base master HEAD)...HEAD. Review every changed file in the branch.

Performing the review

Route each file to a domain via docs-review:references:domain-routing, then apply that domain's criteria plus docs-review:references:shared-criteria. Render the output per docs-review:references:output-format.

For files under content/docs/ or content/blog/, also run Vale and surface its findings per the Style-suggestions render contract in docs-review:references:output-format: entries with "blocker": true (near-zero-false-positive correctness rules) go in 🚨 Outstanding as [style-blocker] bullets; everything else goes under ⚠️ Low-confidence (the **line N:** [style] _category_ — <message> bullet form, grouped under a #### Style suggestions H4, expanded, not counted). Pipe through the categorize filter so the JSON has a deterministic category field — never surface the raw rule name:

vale --no-exit --output=JSON <files> > /tmp/vale-raw.json
python3 .claude/commands/docs-review/scripts/vale-findings-filter.py \
    --in /tmp/vale-raw.json --out /tmp/vale-findings.json
# Render bullets from /tmp/vale-findings.json: use .category, not .rule.

Omit --pr in interactive mode (no diff to intersect; the filter accepts all findings, categorizes, caps). If vale --version fails or vale is not on PATH, skip the Vale step with a one-line note (e.g., "Skipping Vale: not installed. Install via mise install to enable style nits.") and continue the review without Vale findings — don't hard-fail.

For PR-number invocations:

gh pr view {{arg}} --json title,body,files,additions,deletions,labels
gh pr diff {{arg}}

Format for terminal display. Include the scope in the summary, and offer to broaden if useful.

Outcome telemetry (quarterly review)

Every closed PR's pinned review encodes what happened to its findings (✅ Resolved = fixed, concede: = conceded, 🛡️ Disputed … model held = disputed, still-🚨-at-merge = ignored). .claude/commands/docs-review/scripts/scrape-review-outcomes.py derives per-finding outcomes from that structure; the weekly digest (scripts/weekly-digest/digest.py) aggregates the trailing week automatically. Quarterly, run the long-window tuning report:

uv run .claude/commands/docs-review/scripts/scrape-review-outcomes.py \
    --closed-since <quarter-start> --stats

What to do with the numbers: verdict categories with high concede or ignored-at-merge rates are candidates for demoting from the always-🚨 carve-out list (or pruning entirely — edit docs-review:references:output-format §Bucket rules in a PR); recurring prose findings with high fix rates are candidates for promotion into Vale rules so they're caught pre-review. High unconfirmed_at_merge counts mean reviews are routinely stale at merge — evidence for auto-refreshing on trivial deltas. Telemetry informs the tuning; humans make the edits.