Documentation While Fresh

Documentation While Fresh

Document decisions, challenges, and solutions immediately after completing each phase. Captures context that would be lost later, creates invaluable reference for future work.

When to use

• Complex projects with non-obvious design decisions
• Systems with subtle bugs or tricky solutions
• Projects that will be maintained long-term
• When onboarding others to codebase
• Any project taking >1 week
• Open source projects

When NOT to use

• Throwaway prototypes
• Well-understood patterns (standard CRUD app)
• Solo projects you'll never revisit
• When requirements are still in flux

Instructions

Step 1: Create CLAUDE.md at Project Start

Initialize your development log:

# [Project Name] Development Log

## Project Overview
[2-3 sentences: what you're building and why]

## Development Timeline

### Phase 1: [Component Name]
[Will be filled after completing phase]

## Key Patterns Developed
[Will accumulate patterns as discovered]

## Challenges & Solutions
[Will record problems and fixes]

## Performance Results
[Will record measurements]

## Lessons Learned
[Will reflect at end]

## Reusable Components
[Will identify reusable parts]

## Future Improvements
[Will note TODOs and ideas]

## References
[Links to docs, papers, resources used]

Step 2: Document After Each Phase

CRITICAL TIMING: Update CLAUDE.md immediately after completing phase, before starting next phase.

Why: Context is fresh, details are vivid, emotions are real

Template for each phase:

### Phase N: [Component Name]

**Objective**: What were we building?

**Approach**: How did we build it?
1. [Step 1]
2. [Step 2]
3. [Step 3]

**Key Decisions**:
- **Decision 1**: [What we chose]
  - Why: [Rationale]
  - Alternative considered: [What we didn't choose]
  - Trade-off: [What we gained vs lost]

**Challenges & Solutions**:
- **Challenge**: [Problem encountered]
  - Symptom: [How it manifested]
  - Root cause: [Why it happened]
  - Solution: [How we fixed it]
  - Time to fix: [X minutes/hours]
  - Lesson: [What we learned]

**Validation Results**:
- Tests: ✅/❌ [X/Y passing]
- Performance: [Measured vs target]
- Issues found: [Count and brief description]

Step 3: Capture "Why" Not Just "What"

Bad documentation (what):

Used Release/Acquire memory ordering.

Good documentation (why):

**Decision**: Use Release/Acquire memory ordering instead of SeqCst

**Why**: ARM64 has weaker memory model than x86. SeqCst adds unnecessary barriers on ARM, hurting performance. Release/Acquire provides sufficient synchronization for SPSC pattern while being optimally compiled on both architectures.

**Alternative considered**: SeqCst (sequential consistency)
- Pro: Simpler to reason about
- Con: 20-30% slower on ARM64
- Con: Unnecessary barriers

**Trade-off**: Slightly more complex reasoning about memory model, but significantly better performance on target platform (ARM64).

Step 4: Document Challenges in Detail

Template:

**Challenge**: [Brief 1-sentence description]

**Context**: [What we were trying to do when it happened]

**Symptom**: [How the bug manifested]

Compiler error: "attempt to compute 0_usize - 1_usize" Tests failing with wrong output Performance 10x slower than expected

**Root Cause**: [Why it happened]

Rust's default struct layout added unexpected padding CAS loop had no backoff, causing contention Array index calculated after increment instead of before

**Solution**: [How we fixed it]

Reordered struct fields, added explicit padding Added exponential backoff to CAS loop Moved bounds check before array access

**Time to Fix**: [Actual time spent]

**Lesson Learned**: [Generalized insight]

Always use repr(C) for predictable layout Unbounded CAS loops livelock under contention Check bounds before access, not after

**How to Prevent**: [Future proofing]

Add compile-time size assertions Add MAX_RETRIES to all CAS loops Add debug assertions for array access

Step 5: Extract Patterns as You Go

When you solve a problem well, immediately add to "Key Patterns":

## Key Patterns Developed

### Pattern 1: Cache-Line Padded Atomics

**Problem**: False sharing between producer and consumer head/tail

**Solution**:
```rust
#[repr(C, align(64))]
struct CachePadded<T> {
    value: T,
}

When to use: Any atomic shared between threads

Performance impact: 10-100x improvement in contended scenarios

References: ring.rs:15-18

### Step 6: Record Performance Measurements

Don't just say "it's fast", record numbers:

```markdown
## Performance Results

### Ring Buffer (Phase 2)
- **Target**: <50ns per push/pop
- **Measured**: 4.04ns mean (P50)
- **Result**: ✅ 12.5x better than target
- **Outliers**: 4% mild outliers (cache effects)
- **Methodology**: Criterion benchmarks, 1M samples

### Order Book (Phase 3)
- **Target**: <200ns per update
- **Measured**: 2.89-3.50ns mean
- **Result**: ✅ 57-69x better than target
- **CAS retry rate**: <1% (low contention)

Best Practices

✅ Do

• Document immediately - Right after completing phase, not later
• Capture emotions - "This was frustrating because..."
• Record time spent - Helps estimate future work
• Include code snippets - Show actual implementation
• Link to files - Reference specific line numbers
• Record dead ends - What didn't work and why
• Note resources used - Papers, docs, blog posts

❌ Don't

• Don't defer - "I'll document it all at the end" = never happens
• Don't just copy code - Explain the "why"
• Don't sanitize failures - Record real struggles
• Don't skip measurements - "Fast" isn't documentation
• Don't write for yourself only - Write for future you + others

Common Pitfalls

Pitfall 1: Deferring documentation

Problem:

Week 1: Build Phase 1-3
Week 2: Build Phase 4-7
Week 3: "Now I'll document everything"
Week 3 reality: Can't remember why chose X over Y

Solution:

After Phase 1: Document (15 minutes)
After Phase 2: Document (15 minutes)
...
Week 3: Documentation already complete!

Pitfall 2: Documentation without context

Bad:

Fixed bug in bundle.rs

Good:

**Bug**: Array index out of bounds in BundleBuilder::add()

**Symptom**: Panic when adding 16th transaction

**Root cause**: Checked flush condition AFTER incrementing count,
so count reached 16 before checking if >= BUNDLE_MAX (16)

**Solution**: Moved flush check BEFORE array access

**Discovered by**: Benchmark sub-agent during performance testing

**Prevention**: Added debug assertions for index bounds

Pitfall 3: Only documenting successes

Problem: Documentation shows clean path, hides all failures

Solution: Document failed approaches:

**Approaches Tried**:

❌ Attempt 1: SeqCst memory ordering
- Result: 30% slower on ARM64
- Reason failed: Unnecessary memory barriers

❌ Attempt 2: Spinlock for synchronization
- Result: 100x slower, violated lock-free requirement
- Reason failed: Blocking is unacceptable in HFT

✅ Attempt 3: Release/Acquire CAS
- Result: Optimal performance on both ARM and x86
- Why it worked: Minimal synchronization, correct for weak memory model

Integration with Other Skills

With Plan-First Development

Plan defines: WHAT to build
Documentation captures: WHY decisions were made

Example:

• Plan says: "Use Release/Acquire ordering"
• Documentation says: "Chose Release/Acquire because ARM64's weak memory model requires explicit barriers, but SeqCst would add unnecessary overhead..."

With Incremental Validation

Validation: Proves it works
Documentation: Explains how it works

Example:

• Validation: ✅ Tests pass, P99 < 50ns
• Documentation: "Achieved 4ns by using cache-line padding to prevent false sharing..."

Measuring Success

Indicators it's working:

• ✅ Can onboard new developer in 1 hour (read CLAUDE.md)
• ✅ Can revisit project after 6 months and understand decisions
• ✅ Extracted reusable patterns for future projects
• ✅ Documentation took <5% of total project time
• ✅ Others reference your docs as example

Indicators it's failing:

• ❌ "Why did I do it this way?" after 1 week
• ❌ Can't explain design decisions
• ❌ Repeated mistakes in future projects
• ❌ Documentation became burden (took too long)

Related skills

• plan-first-development - Plan provides structure for documentation
• incremental-validation - Validation results go into docs
• parallel-subagent-orchestration - Sub-agents generate complementary docs

Skill Version: 1.0 Last Updated: 2025-01-06