When a pull request, a proposed code change, lands before lunch, choose a diff reader that names the review it performs. A skill is a folder of instructions the model opens when it decides it needs them. A token is the unit your AI plan bills by, about three quarters of a word. 12,820 code review skills load every session, so the choice affects each session before you request a review.
What is in the session before a diff arrives
The context is everything the model can see while it answers you.
The model chooses a skill from its description alone. Vague wording can leave a useful reviewer unused. Of 13,127 skills on this subject, 12,820 load every session and 307 wait to be asked.
Invoking one brings its whole body into the context. That puts the larger cost in the session that needs it. A glob is a filename pattern, like *.tsx. It filters names, but it does not start a review.
What a code review skill costs to keep ready
At 54 tokens, the median preload is closer to a short commit message than a file. It is a small standing cost before you request a review.
Nine in ten cost 125 tokens or less. The largest every-session preload is 565 tokens, in claude-dev-workflow.
If you start 20 sessions a day, the 54-token median becomes 1,080 tokens across a day. A skill that waits to be asked has a median body of 1,146 tokens when loaded. It saves preload while moving cost to review time.
What the scan puts on the line
The scan graded 12,659 skills A, 259 B, 207 C or D, and 2 E or F. An A means the scan found nothing it treats as risky. It does not mean the page lists no finding at all.
The most common findings were 465 network calls, 140 shell commands, 139 reads of agent configuration directories, 124 recursive force deletes, and 50 downloads followed by remote execution.
The triage-contributor-pr skill has a shell-command finding. So does github-code-review. Every finding shows the line that produced it. Inspect that line before making a decision.
Six diff readers to consider
The numbers after each name show preload tokens, file length, and copy status where available. The original is the first copy of a file we saw; everything matching it afterwards is a copy.
- review-pr: a review guide for pull requests in the Deno runtime, with 39 tokens every session, 177 lines, original.
- PR Review: Checks Angular pull requests against package guidance and compatibility. 16 tokens every session, 136 lines, original.
- deerflow-maintainer-orchestrator: Analyses GitHub issues and pull requests, then prepares evidence-based comments, with 88 tokens every session, 294 lines, original.
- review-duplication: Finds duplicated logic and existing utilities during code review, with 37 tokens every session, 70 lines, original, 1 copy.
- pr-status-triage: Investigates CI failures and review comments, with 65 tokens every session, 48 lines, original.
- review: Reviews a pull request or branch diff before merge, with 12 tokens every session, 977 lines, original.
When an event trigger fits better
A hook is a shell command your agent runs for you when something happens.
Choose a hook when timing is the job, such as checking a tool call or another event automatically. A skill cannot be made to fire on an event, which is what a hook is for.
Choose a skill when you want the model to reason about a diff, compatibility, reuse, or review comments. A hook can enforce timing, but it cannot replace that judgement.
Questions people ask
How many tokens does the full review body use? Across these code-review skills, the median body once one is open is 1,336 tokens. That is separate from the 54-token preload, so inspect both before choosing.
What line count should make me hesitate? The heavy examples include a 1,060-line workflow. Use that size as a reason to inspect scope and triggers, not as an automatic rejection.
Does original status matter if a skill has copies? 12,389 entries are originals and 0 are near duplicates. Copy status helps trace provenance, but purpose and preload cost should decide your choice.
Do it now
npx agentmods add <slug> installs one mod by its catalogue slug. Browse code review skills.