perf-review
Skill0 starsChanged 2 months agoApache-2.0View on GitHub
Audit the repo (or a given path, branch, or diff) for performance problems — algorithmic bottlenecks, memory leaks, unnecessary allocation/recomputation, blocking I/O, N+1 access patterns — and report prioritized, evidence-backed findings without changing any code. Use when the user asks to find performance issues, memory leaks, or bottlenecks, or wants to optimize how the app runs or how much memory it uses.
4 other projects use a version of this.
Adopted by
Most recently taken up by
- jpicklyk/task-orchestratortheir own version, first seen Oct 17, 2026
- vhmarquez/argotheir own version, first seen Oct 15, 2026
- octanejs/octanetheir own version, first seen Oct 9, 2026
- DataDog/dd-trace-javatheir own version, first seen Oct 8, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- octanejs/octane★ 1.5ksame name
- DataDog/dd-trace-java★ 739same name
- jpicklyk/task-orchestrator★ 214same name
- vhmarquez/argo★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .cursor/rules/code-quality.mdcCursor rule
- .cursor/rules/git-workflow.mdcCursor rule
- .cursor/rules/hooks.mdcCursor rule
- .cursor/rules/pstack.mdcCursor rule
- verify-seriSkill
- perf-reviewSkill
- thermo-nuclear-code-quality-reviewSkill
The file
We split this skill into the rules it gives an agent
5 rules. 0 are common, 3 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 9
Every finding in this review must trace back to one goal: make the app run faster or use less memory without changing observable behavior.
- Very specific: line 18
No target given → whole repo, default effort medium.
- Distinctive: line 25
low/medium: only high-confidence findings on paths you've confirmed are hot (see below).
- Very specific: line 36
Don't review files uniformly — prioritize code that runs often or on large inputs: entry points and request/event/tool-call handlers, …
- Distinctive: line 96
Do not edit, run, or benchmark anything — this skill only reads code and reports.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Every finding in this review must trace back to one goal: make the app run faster or use less memory without changing observable behavior. A finding that isn't about runtime cost, or whose fix would change behavior, error handling, or output, does not belong here — that's a correctness review, not this one.
Parse the target
Arguments: $ARGUMENTS
- No target given → whole repo, default effort
medium. - A path/directory → scope to that path.
diff, a branch name, or a PR number → scope togit diffagainst the repo's default branch (!git status --shortand `!`git diff --stat` <base>...HEADfor context).- A trailing
low|medium|high|maxtoken sets effort (defaultmedium) — it controls how much you report, not what counts as a real finding:- low/medium: only high-confidence findings on paths you've confirmed are hot (see below). Prefer fewer, certain findings over a long list.
- high/max: broader file coverage and lower-confidence-but-plausible findings are allowed, clearly marked as such.
Exclude generated/vendored code, build output, lockfiles, and binary assets from the scan regardless of scope — grep the manifest/build config to identify these rather than guessing from extension alone.
Find what's actually hot
Don't review files uniformly — prioritize code that runs often or on large inputs: entry points and request/event/tool-call handlers, anything inside a loop that scales with user data, code that runs per-frame/per-tick/per-request/per-message, startup paths (only worth flagging if the delay is user-visible), and anything doing I/O, network calls, subprocess spawning, serialization, or hashing inside a loop.
For each candidate, confirm the finding is real before reporting it: read the call site(s), how often the path actually executes, and the realistic size of the data it handles. An O(n²) loop over an input capped at 5 elements is not a finding — say so to yourself and move on rather than padding the report.
Categories to check
- Algorithmic complexity — nested loops over the same collection, a linear
scan (
find/indexOf/includes/filter) repeated inside an outer loop where a map/set/index would do, repeated recomputation of a value that doesn't change across iterations or calls. - Memory leaks / unbounded growth — listeners, subscriptions, or callbacks registered without a matching removal; caches, maps, or arrays that only grow with no eviction, TTL, or size bound; timers/intervals never cleared; closures that hold a reference to something large well past when it's needed; handles/streams/sockets/ file descriptors/child processes not released on every exit path, including error and early-return paths.
- Blocking work on a hot/shared path — synchronous I/O, crypto, or heavy compute where an async or off-thread alternative exists and the call site can't tolerate blocking (an event loop, a UI thread, a request handler serving concurrent callers).
- Redundant I/O — N+1 query/request patterns, re-reading or re-parsing the same
file/response on every call instead of once, sequential
awaits over independent work that could run concurrently, missing memoization for a pure and expensive function that's called repeatedly with the same inputs. - Unnecessary allocation/copying — deep-cloning or spreading large structures inside a loop or hot path, string concatenation in a loop instead of a builder/join, recreating a stable object/function/regex on every call instead of hoisting it.
- Wasted work — computing or fetching more than the caller needs (a full record to read one field, eager work that could be lazy), work repeated across nearby calls that could be batched or shared.
- Rendering (only if the repo has a UI layer) — unmemoized expensive renders, unstable props/keys forcing full-list re-renders, interleaved layout reads/writes causing thrashing.
Rank and report
Order findings by expected impact: how hot the path is (per-request/frame beats startup-only beats rare/admin-only) times the gap between current and achievable cost. A correct-but-marginal finding on a cold path ranks below a smaller win on a hot one.
For every finding, the suggested fix must preserve behavior. If a plausible fix would also change edge-case handling, error semantics, or output (e.g. trading correctness for speed, dropping a validation step, changing float precision), say so explicitly as part of the finding instead of presenting it as a free win — the user decides whether that tradeoff is acceptable, this review doesn't decide it for them.
Report using the ReportFindings tool, most-severe/highest-impact first:
summary— the mechanism: why this is slow or leaks, not just what the code does.failure_scenario— the concrete condition that makes it manifest (input size, call frequency, time-since-start for a leak) — not "could theoretically be slow."category— one ofalgorithmic-complexity,memory-leak,blocking-io,redundant-io,unnecessary-allocation,wasted-work,rendering.
Do not edit, run, or benchmark anything — this skill only reads code and reports. If the user wants a finding fixed, that's a separate step after they've seen the report.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

