Performance review
Curated, auditable, benchmark-ready Agent Skills library for Claude Code, Codex, OpenCode, Cursor, and more.
npx -y skills add shinzoxD/knackbox --skill performance-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 27 days oldThe repository was created 27 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Diagnose application performance bottlenecks and propose measured fixes. Use whenever the user mentions slow endpoints, high latency, CPU or memory spikes, N+1 queries, profiling results, p95/p99, or asks how to make code or systems faster — even without formal profiler output.
The file declares its own license as Apache-2.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
3.0 KB, as published. Nobody here has run it
Performance Review
Guessing is not a performance strategy. Measure, find the dominant cost, fix that, re-measure. Optimize for the user's SLO (latency, throughput, cost), not micro-benchmark theater.
Workflow
- Define success. Target metric (p95 latency, req/s, memory ceiling), environment (prod-like?), and workload (endpoint, query, batch size).
- Gather evidence. Profiles, traces, EXPLAIN, logs, metrics, or a minimal repro. If missing, prescribe the cheapest measurement first.
- Find the dominant bottleneck. CPU, I/O, lock contention, network chattiness, over-fetch, GC, cold start — pick one primary.
- Propose fixes ordered by expected impact / effort. Prefer algorithmic and I/O reductions before micro-opts.
- Verification plan. How to prove the win and catch regressions.
Common patterns
| Symptom | Often check |
|---|---|
| Slow list endpoint | N+1 queries, missing index, over-serialization |
| High CPU | Hot loops, regex, JSON parse, compression, crypto |
| High memory / OOM | Unbounded caches, loading full tables, leaks |
| Tail latency | Lock contention, GC pauses, cold deps, retries |
| Batch job slow | Sequential I/O, small transactions, no parallelism |
Output format
## Performance review: <surface>
**Goal metric:** …
**Evidence used:** … (or: measurements to collect first)
### Primary bottleneck (hypothesis)
What + why + confidence (low/med/high)
### Recommendations
1. [high impact] … — expected effect, risk, how to verify
2. …
### Do not do yet
Premature micro-opts that distract from the bottleneck.
### Measurement plan
Commands, profilers, or dashboards to confirm.
Rules
- Never claim a speedup percentage without measurement or a clearly labeled estimate.
- One primary bottleneck before a laundry list of nits.
- Correctness and security beat clever caching; call out stale-data risks.
- Distinguish algorithmic O() improvements from constant-factor tweaks.
- If code is not the bottleneck (network, DB plan, noisy neighbor), say so.
- Prefer fixes that add observability (timers, spans) when evidence is thin.
Edge cases
- No profile provided: give a 3-step measurement recipe before deep code rewrites.
- Micro-benchmark only: warn about dead-code elimination and unrealistic inputs; ask for end-to-end numbers.
- Frontend: include bundle size, waterfalls, and main-thread long tasks when relevant — not only server CPU.
- "Make it faster" with no SLO: propose a default (e.g. p95 under X) and proceed.