ct-docs-review

Documentation Review Skill

@skills/_shared/cleo-style-guide.md

Through SDK (preferred)

Read the doc through the docs SSoT — not the filesystem. Reviews that grep loose files miss every doc that lives only as a blob, and miss the version history that makes "what changed?" reviews possible.

Fetch the doc to review

# By slug (preferred — stable handle survives renames):
cleo docs fetch saml-setup-draft

# By attachment ID (when the slug is unknown):
cleo docs fetch att_01HXYZ...

# By SHA-256 (when reviewing a specific version):
cleo docs fetch 7a3f9b...

cleo docs fetch returns the full attachment envelope: metadata, slug, type, owner, refCount, and the bytes (base64-inline for files ≤ 1 MB, storage path for larger blobs).

Diff two versions of the same doc

CLEO doesn't ship a dedicated cleo docs diff verb. Use version listing plus two fetches to compose the diff:

cleo docs versions --for T1234 --name saml-setup-draft
# Returns every SHA-256 version with timestamps.

cleo docs fetch <sha-v1> > /tmp/v1.md
cleo docs fetch <sha-v2> > /tmp/v2.md
diff -u /tmp/v1.md /tmp/v2.md

Reviewers SHOULD anchor every issue to a specific SHA so the author can reproduce the exact state being flagged.

List candidate docs by type

cleo docs list --type spec --project          # all specs in this project
cleo docs list --task T1234 --type research   # research notes attached to T1234

PR review mode — still SDK-grounded

When reviewing a GitHub PR that touches a published doc, fetch the SSoT copy alongside the diff so you can compare the on-disk file in the PR against the canonical blob. Drift between the two is its own review finding — flag it as Issue N: docs SSoT drift.

gh pr diff <pr-number> -- docs/saml-setup.md
cleo docs fetch saml-setup-draft     # SSoT canonical
cleo docs status                     # full drift report

Deprecated: Direct filesystem reads

The legacy "open the .md file in the working tree and review it" pattern is deprecated. The on-disk file may lag the SSoT, the slug- based linkage to the originating task is invisible, and the version history needed for "what changed?" reviews is missing. Migrate to cleo docs fetch <slug> for every review — and run cleo docs status to surface drift before commenting.

Review mode detection

IMPORTANT: Before starting the review, determine which mode to use:

1. PR review mode: If the mcp__github__create_pending_pull_request_review tool is available, you are reviewing a GitHub PR

   • Use the pending review workflow to post all issues as one cohesive review
   • Follow the workflow steps in "PR review mode format" below

2. Local review mode: If the MCP tool is NOT available, output issues in the conversation

   • Format all issues in a numbered markdown list (as described in "Feedback format" below)

Review process

1. Detect review mode - Check if mcp__github__create_pending_pull_request_review is available
2. Read the changes through once to understand intent
3. Check all issues that violate style guide or significantly impact readability
4. Only flag issues worth mentioning - if it won't make a material difference to the reader, skip it
5. REQUIRED: Number ALL feedback sequentially - Start from Issue 1 and increment for each issue found

Review checklist

Run through the diff looking for these issues:

Tone and voice:

• Formal/corporate language ("utilize" not "use", "offerings", etc.)
• "Users" instead of "people" or "companies"
• Excessive exclamation points or overly peppy tone
• Telling readers something is cool instead of showing them

Structure and clarity:

• Important information buried instead of leading
• Verbose text that adds little value
• Paragraphs without clear purpose
• Vague headings that don't convey the point
• Instructions explain "why" before telling "what to do"
• Tasks described as "easy" or "simple"

Links and references:

• Linking the word "here" instead of descriptive text
• Links in headings (unless entire heading is a link)

Formatting:

• Ampersands as "and" substitute (except proper nouns)
• Inconsistent list formatting

Code and examples:

• Code examples that don't work or would error
• Commands not in execution order
• Full-width screenshots instead of scoped UI elements
• Excessive or unnecessary images

Sentence construction:

• Overuse of pronouns when introducing new terms

Quick scan table

Feedback format

MANDATORY REQUIREMENT: Every single issue MUST be numbered sequentially starting from Issue 1.

This numbered format is NON-NEGOTIABLE. It allows users to efficiently reference specific issues (e.g., "fix issues 1, 3, and 5") and track which feedback has been addressed.

Local review mode format

When outputting issues in the conversation (local mode), use this format:

## Issues

**Issue 1: [Brief title]**
Line X: Succinct description of the issue
[code or example]
Suggested fix or succinct explanation

**Issue 2: [Brief title]**
Line Y: Description of the issue
Suggested fix or explanation

**Issue 3: [Brief title]**
...

Examples:

Issue 1: Formal tone Line 15: This could be more conversational. Consider: "You can't..." instead of "You cannot..."

Issue 2: Vague heading Line 8: The heading could be more specific. Try stating the point directly: "Run migrations before upgrading" vs "Upgrade process"

PR review mode format

When posting to GitHub (PR mode), use the pending review workflow:

Workflow steps:

1. Start a review: Use mcp__github__create_pending_pull_request_review to begin a pending review

   • This creates a draft review that won't be visible until submitted

2. Get diff information: Use mcp__github__get_pull_request_diff to understand the code changes and line numbers

   • This helps you determine the correct file paths and line numbers for comments

3. Identify ALL issues: Read through all changes and identify every issue worth mentioning

   • Collect all issues before posting any comments
   • Number them sequentially (Issue 1, Issue 2, Issue 3, etc.)

4. Add review comments: Use mcp__github__add_pull_request_review_comment_to_pending_review for each issue

   • CRITICAL: Post ALL comments in a SINGLE response using multiple tool calls in parallel
   • Each comment should reference a specific file path and line number from the diff
   • Start each comment body with **Issue N: [Brief title]**
   • Include the description and suggested fix

5. Submit the review: Use mcp__github__submit_pending_pull_request_review to publish all comments at once

   • Use event type "COMMENT" (NOT "REQUEST_CHANGES") to make it non-blocking
   • Do NOT include a body message - Leave the body empty or omit it entirely
   • All comments will appear together as one cohesive review

Comment format example:

**Issue 1: Formal tone**

This could be more conversational. Consider: "You can't..." instead of "You cannot..."

IMPORTANT:

• Each issue gets its own review comment attached to the pending review
• Number ALL comments sequentially (Issue 1, Issue 2, Issue 3, etc.)
• Always start the comment body with **Issue N: [Brief title]**
• MUST add all comments in parallel in a single response - Do NOT add them one after another in separate responses
• Do NOT output a summary message to the conversation - only post GitHub review comments
• When submitting the review, do NOT include a body parameter (or leave it empty) to avoid cluttering the PR with summary text
• The review will appear as a single review with multiple comments when submitted

Final check

1. Remove any issues from your assessment that won't make a material difference to the reader if addressed. Only flag issues worth the author's time to fix.
2. Verify all issues are numbered sequentially starting from Issue 1 with no gaps in numbering.
3. Confirm the format exactly matches: **Issue N: [Brief title]** where N is the issue number.
4. In PR mode: Verify each issue was posted as a separate GitHub comment (not output to conversation).

Issue length guidelines

A flood of low-value flags trains the author to ignore reviews. Apply the materiality filter aggressively — be selective up front; flag only what matters.

Workflow discipline

When review should block

The default event is COMMENT (non-blocking). Use REQUEST_CHANGES only when the PR contains issues that, if shipped, would actively harm readers:

• Broken code examples that would fail when copy-pasted
• Outdated security guidance
• Links to deprecated APIs in setup instructions
• Patronizing language to vulnerable audiences

For everything else, use COMMENT. Trust the author to take feedback seriously without being forced.

Re-review after fix

When the author addresses feedback and pushes new commits, run the review again on the updated diff. Don't re-flag issues already fixed; do flag new issues introduced in the fix.

For re-reviews, start the comment with a status line:

**Issue 2 (UPDATED): Vague heading**

The previous fix changed "Setup" to "Setup Steps" — still vague.
Try "Install dependencies, then run init".

The "(UPDATED)" tag signals iteration, not a brand-new flag.

See references/

Progressive disclosure — load on demand only:

• references/style-violations.md — taxonomy of violations with detection patterns, issue titles, and remediation messages
• references/pr-review-mode.md — pending-review workflow, comment format, gh CLI fallback, materiality filter
• references/inline-comment-patterns.md — comment anatomy, single-line/multi-line/pattern fixes, anti-patterns