code-review-digest-writer

Code Review Digest Writer Skill

When to Use This Skill

• You want a weekly (or custom window) code-review digest for the current

repository or project, based on PR review comments.

• You have a start date and end date and want a markdown newsletter

summarizing what was taught in PR review feedback in that period.

• You want to highlight themes, repeated issues, and concrete best practices

rather than just listing PRs.

If the user does not provide both a start and end date, ask them to specify: YYYY-MM-DD → YYYY-MM-DD before proceeding.

Example Prompts

• “Generate a code review digest for this repo for 2025-02-01 → 2025-02-14.”
• “Create a weekly PR review digest for 2025-03-10 → 2025-03-17, using any existing digests as historical context.”
• “From 2025-04-01 to 2025-04-30, summarize what reviewers focused on in owner/service-repo and highlight repeated issues.”
• “Write a newsletter-style code review digest for this project for 2025-05-15 → 2025-05-29, tagging themes as [NEW] or [REPEAT].”

Scope & Repositories

• Target repo: whichever project you want a digest for (typically the repo you

currently have open in Claude Code).

• Ideally the repo contains docs/review-digests/AGENTS.md with

project-specific digest guidelines. If not, fall back to the generic structure described in this Skill and any existing digest files.

• This Skill must only modify docs under:
• docs/review-digests/YYYY-MM-DD.md

where the date is the end date of the reporting window.

• Do not modify application code, tests, or configs in the target repo while

this Skill is active. All changes should be within docs/review-digests/.

• If docs/review-digests/ does not exist yet, create the directory before

writing the digest file, so future digests can be added and past ones read.

Required Local Tools & Assumptions

When using Bash, assume:

• gh (GitHub CLI) is installed and authenticated for the target repository

(or for a default GitHub identity that can see it).

• Current working directory is the target repo root (or pass --repo owner/name

explicitly to gh commands, if needed).

If gh is not available, gracefully fall back to:

• Summarizing based on locally available PR notes or docs, and
• Clearly stating in the digest that it was generated with partial data.

High-Level Workflow

When asked to generate or update a digest, follow this workflow:

1. Confirm time window
2. Ensure you have start_date and end_date (inclusive).
3. Confirm with the user if there is any ambiguity.

1. Load local digest guidelines
2. If it exists, open docs/review-digests/AGENTS.md and read it carefully.
3. When present, treat that file as the source of truth for:
4. What the digest is.
5. Where it should be written.
6. Required structure and link style.
7. How to detect and label repeated issues.
8. If it does not exist, follow the layout described later in this Skill and

use any existing digest files in docs/review-digests/ as a reference.

1. Inspect existing digests
2. Ensure the docs/review-digests/ directory exists:
3. If it does not, create it; in that case there will be no past digests yet.
4. Use Glob or Bash to list docs/review-digests/*.md.
5. Load at least the last 3–4 digests (if present).
6. Extract their themes and repeated issues (e.g., fixture reuse,

blind-index invariants, Query() regression patterns, etc.).

• You will use these to detect when issues are recurring.

1. Fetch PR and review data for the window
2. Use Bash with gh to query PRs whose:
3. createdAt is between [start_date, end_date], OR
4. closedAt is between [start_date, end_date].
5. Deduplicate PR numbers.
6. For each selected PR:
7. Fetch top-level comments (PR discussion).
8. Fetch review-thread comments with code context (via GraphQL).
9. Prefer comments from:
10. Human reviewers (__typename == "User").
11. AI reviewers that contain substantial review content

(e.g., Claude, Copilot PR reviewer).

• Exclude noisy infrastructure/bot comments with no review content

(e.g., github-actions, log-only bots, CI status updates).

1. Cluster feedback into themes
2. Read comments and diff context enough to understand:
3. What behavior or pattern was being discussed.
4. What best practice or correction was suggested.
5. Group comments across PRs into themes, such as:
6. Logging, Sentry, and performance instrumentation.
7. Tests, fixtures, and code structure.
8. Security, access control, and PII handling.
9. Domain-specific design and invariants for this repository.
10. Migrations & tooling.
11. Process and meta-patterns in reviews.

1. Detect repeated issues
2. For each current theme, compare it conceptually to themes you extracted

from previous digests (step 3).

• If the same pattern appears again (e.g., “use TypedDict instead of

dict[str, Any] in payloads”, “avoid Django Ninja Query() constants”, “reuse shared fixtures instead of copy-paste”), treat it as a repeated issue.

• Use labels:
• [NEW] for themes that appear for the first time.
• [REPEAT] for themes that have appeared in previous digests.

1. Draft the digest file
2. Target path: docs/review-digests/END_DATE.md
3. Example: period 2025-11-13 → 2025-11-27 → docs/review-digests/2025-11-27.md.
4. Follow the layout described in docs/review-digests/AGENTS.md and the

most recent digest, including:

• Title with repo and period.
• Overview section with 3–6 bullets summarizing main themes.
• Thematic sections (numbered) that group related feedback.
• A closing section (e.g., “How to Use This Digest”).
• Within each section:
• Explain the practice in plain language.
• Include 1–3 concrete, generalized examples.
• Call out whether this is [NEW] or [REPEAT].
• Emphasize the “why” (business impact, correctness, safety, DX).

1. Linking to PRs and comments
2. In the body of the digest, use reference-style links only:
3. [#2519 – Fix Teams Start Survey race condition][pr-2519]
4. Key comment: [Fixture reuse recommendation][c-2519-3]
5. At the bottom of the file, define every link once:
6. pr-<number> for PRs.
7. c-<number>-<n> for specific review comments.
8. ic-<number>-<n> for issue comments, if needed.
9. Reuse identifiers consistently when the same comment is referenced in

multiple sections.

1. Respect tone and intent
2. The digest is a newsletter, not a blame report.
3. Highlight:
4. What the team is learning.
5. Where we’re improving.
6. Where patterns are still repeating and need attention.
7. Make guidance actionable (e.g., “When adding a new CSV mapping endpoint,

always run through the project’s PII and security checklist docs.”).

1. Save and review
2. Use Edit to create or update the digest file for END_DATE.
3. Re-open the file after writing to sanity-check:
4. Structure matches the prior digests.
5. Links resolve correctly and have definitions at the bottom.
6. [NEW] / [REPEAT] tags are applied consistently.
7. No accidental code changes occurred in the repo.

Output Expectations

When this Skill is active and asked to generate a digest, your final output should be:

• A single markdown file under docs/review-digests/YYYY-MM-DD.md.
• A short natural-language summary back to the user describing:
• The period covered.
• The main themes identified.
• How many themes were [REPEAT] vs [NEW].

If you were unable to access GitHub or some PRs, clearly note in the digest and in your summary which data sources were missing and how that might limit the digest.

Severity / Emphasis Tags

Instead of issue severities,