Rob Pike's 5 Rules of Programming
The Rules
- You can't tell where a program is going to spend its time. Bottlenecks occur in surprising places. Don't guess — prove it.
- Measure. Don't tune for speed until you've measured. Even then, don't unless one part of the code overwhelms the rest.
- Fancy algorithms are slow when n is small, and n is usually small. Big-O doesn't matter when constants dominate. Use Rule 2 first.
- Fancy algorithms are buggier than simple ones. Use simple algorithms and simple data structures.
- Data dominates. Choose the right data structures and the algorithms become self-evident. "Write stupid code that uses smart objects."
How to Apply
Before Any Optimization
Step 0: Check for Existing Instrumentation
Before asking "have you measured?", determine whether measurement is even possible right now.
Scan the codebase for signs of existing instrumentation:
- Logging: look for logger imports, log calls, structured logging libraries
- Profiling: look for profiler imports, benchmark files, tracing setup
- Timing: look for duration measurements, stopwatch patterns, timing decorators
- APM/Observability: look for metrics exports, spans, trace contexts
Then ask the user:
- If instrumentation exists: "I found logging/profiling in [locations]. Are there specific areas you suspect are slow, or should we look at what the existing measurements tell us?"
- If instrumentation is missing or sparse: "There's no measurement in place to prove where time is being spent. Before optimizing anything — where do you suspect the bottleneck is? Let's add measurement there first, then let the data decide."
The goal is NOT to prescribe a specific tool — Claude already knows the right profiling approach for the language. The goal is to make sure measurement exists before any optimization conversation continues. If there is nothing to measure with, the first action is adding instrumentation, not changing code.
Step 1: Ask the Measurement Questions
Stop and ask these questions in order:
- "Have I measured?" — If no, measure first. Any optimization without measurement data is premature. Use whatever profiling tool is natural for the project's language and ecosystem.
- "Does one part overwhelm the rest?" — If no single area dominates, there is nothing worth optimizing. Small improvements spread across many areas rarely matter.
- "What's n?" — If n is small (and it usually is), the simple O(n²) approach likely beats the clever O(n log n) one due to constants, cache behavior, and implementation complexity.
- "Is this a data structure problem?" — Before changing the algorithm, consider whether a different data structure makes the problem trivial. The right structure often eliminates the need for a clever algorithm entirely.
- "Is the added complexity worth it?" — Simple code that is 10% slower is almost always preferable to clever code that is fragile and hard to maintain.
Anti-Patterns to Block
When you catch yourself or the user doing any of these, STOP and redirect:
| Impulse |
Rule violated |
Response |
| "This loop looks slow, let me optimize it" |
Rule 1 |
Have you profiled? The bottleneck may be elsewhere entirely. |
| "Let me add a cache here" |
Rule 2 |
Measure first. Does this path actually dominate runtime? |
| "Let me use a B-tree / trie / skip list" |
Rule 3 |
What's n? If small, a sorted slice + binary search wins. |
| "Let me implement a custom allocator" |
Rule 4 |
Start simple. Measure. Only get fancy if data forces you. |
| "The algorithm is O(n²), needs fixing" |
Rule 3 |
What's n? O(n²) with n=100 is 10μs. Measure first. |
| "Let me parallelize this" |
Rule 2 |
Is this actually CPU-bound? Measure. Often it's I/O. |
When Optimization IS Justified
Proceed with optimization only when ALL of these are true:
- You have measurement data showing a specific bottleneck
- That bottleneck dominates overall runtime (not just 5-10% of it)
- The proposed fix is the simplest change that addresses the measured problem
- You will re-measure after the change to confirm improvement
1---2name: agentic-rob-pike3description: Rob Pike's 5 Rules of Programming — a decision framework that prevents premature optimization and enforces measurement-driven development. Use when the user says "optimize", "slow", "performance", "bottleneck", "speed up", "make faster", "too slow", or any request to improve code speed/efficiency. Also use when you notice yourself about to suggest a performance optimization without measurement data. This is a thinking discipline, not a tooling workflow.4---56# Rob Pike's 5 Rules of Programming78## The Rules9101. **You can't tell where a program is going to spend its time.** Bottlenecks occur in surprising places. Don't guess — prove it.112. **Measure.** Don't tune for speed until you've measured. Even then, don't unless one part of the code overwhelms the rest.123. **Fancy algorithms are slow when n is small, and n is usually small.** Big-O doesn't matter when constants dominate. Use Rule 2 first.134. **Fancy algorithms are buggier than simple ones.** Use simple algorithms and simple data structures.145. **Data dominates.** Choose the right data structures and the algorithms become self-evident. "Write stupid code that uses smart objects."1516## How to Apply1718### Before Any Optimization1920#### Step 0: Check for Existing Instrumentation2122Before asking "have you measured?", determine whether measurement is even **possible** right now.2324**Scan the codebase** for signs of existing instrumentation:25- Logging: look for logger imports, log calls, structured logging libraries26- Profiling: look for profiler imports, benchmark files, tracing setup27- Timing: look for duration measurements, stopwatch patterns, timing decorators28- APM/Observability: look for metrics exports, spans, trace contexts2930**Then ask the user:**31321. If instrumentation **exists**: "I found logging/profiling in [locations]. Are there specific areas you suspect are slow, or should we look at what the existing measurements tell us?"332. If instrumentation is **missing or sparse**: "There's no measurement in place to prove where time is being spent. Before optimizing anything — where do you suspect the bottleneck is? Let's add measurement there first, then let the data decide."3435The goal is NOT to prescribe a specific tool — Claude already knows the right profiling approach for the language. The goal is to **make sure measurement exists before any optimization conversation continues.** If there is nothing to measure with, the first action is adding instrumentation, not changing code.3637#### Step 1: Ask the Measurement Questions3839Stop and ask these questions in order:40411. **"Have I measured?"** — If no, measure first. Any optimization without measurement data is premature. Use whatever profiling tool is natural for the project's language and ecosystem.422. **"Does one part overwhelm the rest?"** — If no single area dominates, there is nothing worth optimizing. Small improvements spread across many areas rarely matter.433. **"What's n?"** — If n is small (and it usually is), the simple O(n²) approach likely beats the clever O(n log n) one due to constants, cache behavior, and implementation complexity.444. **"Is this a data structure problem?"** — Before changing the algorithm, consider whether a different data structure makes the problem trivial. The right structure often eliminates the need for a clever algorithm entirely.455. **"Is the added complexity worth it?"** — Simple code that is 10% slower is almost always preferable to clever code that is fragile and hard to maintain.4647### Anti-Patterns to Block4849When you catch yourself or the user doing any of these, STOP and redirect:5051| Impulse | Rule violated | Response |52|---|---|---|53| "This loop looks slow, let me optimize it" | Rule 1 | Have you profiled? The bottleneck may be elsewhere entirely. |54| "Let me add a cache here" | Rule 2 | Measure first. Does this path actually dominate runtime? |55| "Let me use a B-tree / trie / skip list" | Rule 3 | What's n? If small, a sorted slice + binary search wins. |56| "Let me implement a custom allocator" | Rule 4 | Start simple. Measure. Only get fancy if data forces you. |57| "The algorithm is O(n²), needs fixing" | Rule 3 | What's n? O(n²) with n=100 is 10μs. Measure first. |58| "Let me parallelize this" | Rule 2 | Is this actually CPU-bound? Measure. Often it's I/O. |5960### When Optimization IS Justified6162Proceed with optimization only when ALL of these are true:6364- You have measurement data showing a specific bottleneck65- That bottleneck dominates overall runtime (not just 5-10% of it)66- The proposed fix is the simplest change that addresses the measured problem67- You will re-measure after the change to confirm improvement
tmdgusya