SWE Performance Hunt

What This Skill Does

Use this skill to identify performance bottlenecks conservatively and turn them into a small, reviewable backlog.

The goal is to separate:

• measured bottlenecks
• high-confidence hotspots supported by repo evidence
• weak suspicions that do not yet justify optimization work

The output should be a ranked set of bottlenecks plus the smallest next experiment or fix for each item that clears the evidence bar.

When To Use

Use this skill when the user wants to:

• find performance bottlenecks in a service, job, page, query path, or repo
• understand why a specific flow is slow
• audit recent performance regressions with concrete evidence
• rank hot paths before starting optimization work
• run a recurring performance review over time

Do Not Use

Do not use this skill for:

• live incident response or active outage triage
• generic observability coverage audits
• broad cleanup or refactor planning with no performance target
• speculative micro-optimization with no measured user or system impact
• architecture rewrites that are not justified by bottleneck evidence

Inputs To Confirm

Confirm or infer:

• repository, service, or package scope
• target flow, workload, or endpoint
• primary metric of interest, such as latency, throughput, memory, startup, or bundle size
• available evidence sources such as profiles, traces, benchmarks, or query plans
• whether the user wants a report-only pass or a small optimization backlog
• any guardrails around risky changes or expensive tests

If the request is too broad, narrow it to one flow and one metric before doing deep analysis.

Evidence Sources

Strong evidence includes:

• profiler output or flamegraphs
• targeted benchmarks
• load-test results
• query plans or database timing data
• frontend bundle analysis or render timing output
• traces or metrics tied to the scoped flow
• recent regression data from CI or performance tests

Useful repo evidence includes:

• obviously expensive loops or repeated work on hot paths
• missing caches or invalidation patterns
• N+1 queries or repeated remote calls
• unnecessary serialization, parsing, or data copying
• oversized client bundles or repeated rerender triggers
• synchronization or locking patterns likely to throttle throughput

If runtime data is absent, use repo evidence carefully and label findings as inferred until measured.

Parallelization Rule

Use parallel work only for clearly disjoint surfaces, such as separate services, pages, or jobs.

• keep each session on one performance surface
• avoid splitting shared libraries or overlapping hot paths across sessions
• return raw evidence from each session, not just conclusions
• rank the combined backlog centrally

If surfaces overlap heavily, keep the hunt serial and smaller.

Instructions

Step 1: Define The Performance Surface

Lock the scope to the smallest surface that matches the request, such as:

• one endpoint
• one background job
• one page load
• one query path
• one benchmark suite

Name the primary metric and why it matters.

Step 2: Gather The Strongest Evidence First

Start from measured evidence when available.

Prefer:

• existing benchmark outputs
• profiler traces
• load-test artifacts
• query plans
• frontend performance reports

Do not lead with optimization folklore when stronger evidence exists.

Step 3: Map The Critical Path

Trace the work that dominates the scoped flow:

• entry points
• repeated loops or render cycles
• database and network boundaries
• serialization and parsing work
• cache lookups and misses
• synchronous chokepoints

Identify where time, memory, or bytes are likely being spent.

Step 4: Separate Bottlenecks From Suspicion

Bucket findings into:

• strong evidence: measured bottleneck or direct regression signal
• moderate evidence: repo-backed hotspot with clear likely impact
• weak evidence: plausible issue that still needs measurement

Only recommend fixes or experiments for strong and moderate findings.

Step 5: Rank By User And System Impact

Prioritize bottlenecks that:

• materially affect the target metric
• sit on a critical or frequent path
• have a small, low-risk next step
• can be verified with narrow checks

Avoid broad rewrites when a smaller validation step would answer the question.

Step 6: Propose Minimal Follow-Up Work

For each bottleneck that clears the bar, propose:

• the affected surface
• the evidence
• the likely cause
• the smallest practical experiment or fix
• the validation command, benchmark, or measurement to rerun

Prefer measurement-tight loops over broad optimization programs.

Step 7: Call Out Unknowns

Be explicit about:

• missing runtime data
• surfaces you could not verify
• findings that are currently inferred rather than measured

If the data is too weak, say so and stop at the measurement plan.

Output Requirements

Provide a report with these sections:

1. Scope and target metric
2. Evidence reviewed
3. Ranked bottlenecks
4. Weak-signal suspicions skipped
5. Proposed follow-up experiments or fixes
6. Unknowns or limits

For each ranked bottleneck, include:

• surface
• target metric or regression
• concrete evidence
• likely cause
• smallest next step
• validation path
• confidence

For each weak-signal item, include:

• surface
• what looked suspicious
• why the evidence is insufficient
• explicit no-fix decision

If there are no credible bottlenecks, say so plainly and explain what evidence looked healthy.

Quality Bar

• Stay tied to the scoped flow and metric.
• Prefer measured evidence over performance folklore.
• Distinguish observed bottlenecks from inferred hotspots honestly.
• Recommend the smallest validation-tight next step.
• Do not drift into generic cleanup, observability, or architecture work.