Hard Rules

• Never optimize code you haven't profiled. Intuition about what's slow is wrong more often than it's right. Measure first, every time.
• Verify correctness after every optimization. Run the full test suite and confirm output matches the original within acceptable tolerance. A faster wrong answer is still wrong.
• Record before/after numbers. No "I think it's faster now." If you can't show a measured delta, the optimization isn't done.

Performance Tuning

Measure first. Optimize second. Verify improvement with data.

The diagnostic workflow is: measure → identify the bottleneck → optimize → benchmark. If your host also exposes a profiling-tool skill (cProfile flags, memory profiler setup), use it alongside this workflow — this skill owns the method, that owns the tool syntax.

Before Optimizing: Check if You Should

Premature optimization checklist — if ANY of these are true, stop and reconsider:

• [ ] The code runs fast enough for current usage (users aren't waiting)
• [ ] This code runs rarely (not in a hot path)
• [ ] You haven't profiled to confirm this is the bottleneck
• [ ] There's a simpler architectural fix (e.g., cache the result, reduce API calls)
• [ ] The code is hard to understand and optimizing will make it worse

Only proceed when: (1) you have a measured bottleneck and (2) it actually impacts the user.

Step 1: Measure (Profile First)

Python — cProfile

# Profile a script
python -m cProfile -s cumtime your_script.py

# Profile with output to file (for large reports)
python -m cProfile -o profile.out your_script.py
python -c "import pstats; p = pstats.Stats('profile.out'); p.sort_stats('cumulative'); p.print_stats(20)"

Python — memory_profiler

pip install memory_profiler

# Add @profile decorator to the function
# Then run:
python -m memory_profiler your_script.py

Python — line_profiler (line-by-line)

pip install line_profiler

# Add @profile decorator to the function
# Then run:
kernprof -l -v your_script.py

Node.js

# Built-in profiler
node --prof your_script.js

# Analyze output
node --prof-process isolate-*.log > profile.txt

Step 2: Read the Profile Output

What to look for:

• cumtime (Python) — total time spent in function including callees → find the top offenders
• tottime (Python) — time in function excluding callees → find the actual slow code
• High call counts on unexpected functions → loop-in-loop or unnecessary repeated work

Hot-loop amplification: Look for functions called thousands of times inside a tight loop (per frame, per row, per request, per event). One slow function called 10,000 times = 10,000x the pain. Pull invariant work out of the loop.

Step 3: Optimize the Bottleneck

Python — Common Patterns

Vectorize with pandas/numpy instead of loops:

# SLOW: Python loop over DataFrame
for i, row in df.iterrows():
    result.append(row['price'] * row['qty'])

# FAST: Vectorized operation
result = df['price'] * df['qty']

Chunk large DataFrames:

# Process in chunks instead of loading all data at once
for chunk in pd.read_csv('large_file.csv', chunksize=10000):
    process(chunk)

Use sets for membership testing:

# SLOW: O(n) list lookup
if item in my_list:  # searches entire list

# FAST: O(1) set lookup
if item in my_set:

Cache expensive computations:

from functools import lru_cache

@lru_cache(maxsize=128)
def expensive_calc(n: int) -> float:
    ...  # computed once, cached after

Hoist invariant work out of hot loops

A common bottleneck in any tick-by-tick or row-by-row pipeline: recomputing a derived series on every iteration when it could be computed once before the loop.

# SLOW: recompute the same indicator on every iteration
for i, row in df.iterrows():
    signal = compute_metric_for_one_row(row)

# FAST: pre-compute the entire series once, vectorized
df['metric'] = compute_metric_series(df['value'])  # once, vectorized

Database / SQL query optimization

# Add an index on frequently filtered columns
conn.execute("CREATE INDEX IF NOT EXISTS idx_entity_date ON records (entity_id, record_date)")

# Inspect the query plan to see what the engine is doing
conn.execute("EXPLAIN QUERY PLAN SELECT ...")  # SQLite; use EXPLAIN ANALYZE on Postgres

Avoid per-row Python callbacks

# SLOW: apply() calls a Python function per row
df['result'] = df.apply(lambda row: complex_calc(row), axis=1)

# FAST: vectorized numpy operation
df['result'] = np.where(df['value'] > threshold, df['a'], df['b'])

Step 4: Benchmark Before and After

Always record results. No "I think it's faster now."

## Performance Benchmark — [Feature/Script Name]
Date: YYYY-MM-DD

| Metric | Before | After | Improvement |
|--------|--------|-------|-------------|
| Job runtime (full dataset) | 45s | 8s | 82% faster |
| Memory peak | 2.1 GB | 340 MB | 84% less |
| Rows processed/sec | 1,200 | 8,400 | 7x faster |

Commit this table alongside the change so the gain is part of the record, not folklore.

Quick Reference

Trigger Conditions

Use this skill when:

• Code works correctly but runs too slowly, hangs, or "takes forever"
• A pipeline, batch job, or simulation has a bottleneck
• Memory usage is too high or grows over a long-running process
• Latency on a hot path needs to come down
• You need a before/after benchmark to justify an optimization

Phrases that trigger: "this is slow", "optimize", "profile this", "pipeline bottleneck", "memory usage", "benchmark", "it's hanging", "needs to be faster", "latency", "why is this slow".

Out of Scope

• NOT for fixing bugs that produce wrong output — run a systematic diagnosis (use debug-session) for correctness issues; this skill is for code that works correctly but runs too slowly.
• NOT for refactoring code structure without a measured bottleneck — use refactor-session for structural improvements; this skill requires a profiled performance problem.
• NOT for slowness caused by missing or stale data rather than code — diagnose the data path first (use debug-session) if the problem isn't the code itself.

Common Traps

• Optimizing without profiling first: Intuition about what's slow is wrong more often than right. Spending hours optimizing a function that accounts for 2% of runtime while the real bottleneck (database query, network call, serialization) goes untouched. Always profile before touching code.
• Micro-benchmarks that don't reflect real workloads: A function benchmarked in isolation with 100 items may behave completely differently with 1M items (cache misses, memory pressure, GC pauses). Benchmark with production-representative data sizes and shapes.
• Optimization that breaks correctness: Replacing a stable sort with an unstable one, caching a result that should be recomputed, or removing a "redundant" check that handled an edge case. Always run the full test suite after any optimization and verify output matches exactly.
• Caching without an invalidation strategy: Adding @lru_cache or a manual cache speeds things up until stale data causes wrong results. Every cache needs a clear invalidation policy — time-based, event-based, or size-based. Unbounded caches also leak memory over long-running processes.
• Vectorization that changes floating-point results: NumPy/pandas vectorized operations may use a different floating-point evaluation order than Python loops, producing slightly different results. For financial calculations or ML model reproducibility, verify that optimized output matches the original within acceptable tolerance.

Skill Chain