comark ui
Esc

Agent skill

Review document

View as MarkdownOpen in ChatGPTOpen in Claude

Your agent writes the review a reader can act on in a minute, with the evidence that makes each finding hold up.

Install the global Vite+ CLI before running this command. Without it, replace vp dlx with pnpm dlx.

Pick review-document from the list.

The skill's example

Playground

Review: avatar cache

Request changes

One blocking finding: after a large avatar loads, every smaller one on the page shows the large image. The fix is one line.

Reviewed acme/web#418 at 9f3c2e1.

1. Small avatars load the 512px image (blocking)

Section titled "1. Small avatars load the 512px image (blocking)"

src/avatars/cache.ts:18. The cache key is the user's id alone, so once the profile header loads someone's 512px avatar, the 32px avatars in comments get the same file. Reproduced with a test:

pnpm test src/avatars1 passed1 failed180 ms

cache.test.ts › returns a 32px avatar after a 512px one

expected 512 to be 32

AssertionError: expected 512 to be 32
      at cache.test.ts:27:31
Fixsrc/avatars/cache.ts+1 −1
const key = user.id;
const key = `${user.id}:${size}`;

2. The cache never evicts (should fix)

Section titled "2. The cache never evicts (should fix)"

src/avatars/cache.ts:9. Read in the code, not measured: the map grows by one entry per user and size for as long as the tab stays open, which on the team directory is thousands of images.

const cache = new Map<string, Blob>();
src/avatars/cache.ts

Cap it, such as an LRU of 500 entries, or decide the tab's lifetime is short enough not to.

Checked

Section titled "Checked"
  • Ran pnpm test src/avatars: one failure, above.
  • Opened a profile on the preview at 1440px and watched the comment avatars after the header loaded.

Not checked

Section titled "Not checked"
  • Server rendering: the cache only runs in the browser, and I didn't confirm the server skips it.
  • Safari's memory use.

In the skill folder

What it does

Section titled "What it does"

The skill writes up a finished review, with a verdict first and then only the findings that change it, each with its location, evidence, and fix. It doesn't review anything. Run your usual review first, and the skill turns its findings into a page a reader acts on in a minute.

When to use it

Section titled "When to use it"

Ask once the review is done, such as "Write up your review of #418." Your agent also loads it for a design review, a QA pass, or an audit. To be sure, say "Use the review-document skill."

Choose by what you have:

  • Findings from a code, design, or QA review, and a reader who must decide: this skill.
  • A diff nobody has reviewed yet: your review process first, then this skill.
  • One visible change to prove: Before/after evidence alone. Inside a review, its compare is one finding's evidence.

It's working if

Section titled "It's working if"
  • The callout on top says whether to merge and names the finding behind it.
  • Every finding has a rank, blocking first, and a title that says what goes wrong and for whom.
  • Each finding says whether the agent reproduced it or read it in the code, so you know which to check yourself.
  • The page ends with what the agent checked and what it didn't.
  • No summary repeats the findings, and there's no list of what's good unless you asked for one.