Skip to content
Filip Stack
Documentation: pr-review-churn

pr-review-churn

Analyzes recurring human review objections on GitHub PRs and proposes guidance changes.

Install
On its own
Runs
Picked when a request matches
Source
SKILL.md
npx skills add filipgutica/filip-stack -g -a claude-code codex -s pr-review-churn -y

What it does

Human review threads provide the evidence: the agent groups comments that object to the same engineering behavior, counts how many PRs each pattern spans, and preserves examples. Proposed guidance comes from repeated objections rather than one isolated comment.

When to use it

Install the skill with the command above. You also need Python 3.9 or later and an authenticated GitHub CLI, gh.

Name the repositories and lookback period in your request. The agent selects the skill for review-churn analysis. For example:

Analyze review churn on acme/web and acme/api for the last 90 days.

Neither repositories nor lookback has a default. If either is missing, the agent asks and waits before collecting PRs. It uses your GitHub login as the author unless you name someone else.

Use the report to find recurring review objections and choose where team guidance might prevent them.

How it works

A bundled collector fetches review threads with file locations and resolved or outdated status. Its cohort is PRs created on or after the cutoff date. Reviews submitted during that period on older PRs are excluded; the report states that limit.

The agent filters bot-only comments, resolved typo nits, and author comments that do not clarify another reviewer's point. Useful human feedback stays even when a thread is resolved or outdated. Comments join a pattern only when they object to the same behavior.

A failed repository does not stop the others. The report identifies failed or partial coverage rather than treating a successful collector exit as proof of completeness.

SKILL.md has the collector and report contract.

What you get

  • A Markdown report with coverage limits, pattern counts, PR links, and short examples.
  • Proposed guidance changes aimed at existing files, such as AGENTS.md or a review checklist, and recommended next actions.
  • A distinction between fully covered, partial, and inaccessible repositories.

The analysis is read-only: it posts no comments and changes no repository, skill, or guidance file. Request a follow-up implementation to apply proposals; technical-writing can help write the guidance.

  • review

    Reviews one exact change or document and reports only the findings it can back with evidence.

    Workflow
  • technical-writing

    Writes or revises technical documents so readers can act on them without guessing.

    Workflow