Lessons Learned

Iron Law

NO PHASE SHIPS WITHOUT LESSON CAPTURE

No exceptions:

• Not for "small phases" or "just a bugfix"
• Not for "we're in a hurry"
• Not for "nothing notable happened"
• Even smooth phases confirm what's working — capture it

When to Use

• After phase completion during /shipyard:ship (Step 3a)
• When reflecting on completed work to extract reusable knowledge
• When a build summary contains notable discoveries worth preserving

Natural Language Triggers

• "what did we learn", "capture lessons", "retrospective", "lessons learned"

When NOT to Use

• Mid-implementation (wait until the phase ships — partial lessons lack full context)
• Documenting code design decisions (put those in inline comments or architecture docs)
• Tracking active bugs (use issue tracker; lessons capture completed learning, not open problems)

Timing Rule

Run immediately after ship, before the next phase starts. Lessons captured 24+ hours later lose the "why" behind decisions. If you're about to start the next phase, capture lessons first.

Overview

The lessons-learned system captures discoveries, patterns, and pitfalls found during implementation and feeds them back into the project. Lessons are stored in .shipyard/LESSONS.md and optionally surfaced in CLAUDE.md so future agents benefit from past experience.

LESSONS.md Format

Store lessons in .shipyard/LESSONS.md using this exact structure:

# Shipyard Lessons Learned

## [YYYY-MM-DD] Phase N: {Phase Name}

### What Went Well
- {Bullet point}

### Surprises / Discoveries
- {Pattern discovered}

### Pitfalls to Avoid
- {Anti-pattern encountered}

### Process Improvements
- {Workflow enhancement}

---

New entries are prepended after the # Shipyard Lessons Learned heading so the most recent phase appears first. Each phase gets its own dated section with all four subsections.

Prioritization Order

When multiple lessons compete for a limited CLAUDE.md budget, prioritize in this order:

1. Pitfalls to Avoid — prevents repeated mistakes (highest value)
2. Process Improvements — improves next phase execution
3. Surprises / Discoveries — corrects wrong mental models
4. What Went Well — confirms what to repeat (lowest urgency — already working)

Structured Prompts

Present these four questions to the user during lesson capture:

1. What went well in this phase? -- Patterns, tools, or approaches that worked effectively.
2. What surprised you or what did you learn? -- Unexpected behaviors, new techniques, or revised assumptions.
3. What should future work avoid? -- Anti-patterns, dead ends, or approaches that caused problems.
4. Any process improvements discovered? -- Workflow changes, tooling suggestions, or efficiency gains.

Pre-populate suggested answers from build artifacts before asking (see Pre-Population below).

Pre-Population

Before presenting prompts, extract candidate lessons from completed build summaries:

1. Read all SUMMARY-*.md files in .shipyard/phases/{N}/results/.
2. Extract entries from "Issues Encountered" sections -- these often contain workarounds and edge cases.
3. Extract entries from "Decisions Made" sections -- these capture rationale worth preserving.
4. Present extracted items as pre-populated suggestions the user can accept, edit, or discard.

This reduces friction and ensures discoveries documented during building are not lost.

Multi-Agent Synthesis

When lessons come from multiple builder agents working in parallel:

1. Collect — gather SUMMARY-*.md from all agents in the phase
2. Group — sort entries by the four lesson categories
3. De-duplicate — merge entries that describe the same issue with different wording
4. Prioritize by frequency — if 3 agents hit the same pitfall, it ranks higher than a one-off
5. Cross-reference — check against existing LESSONS.md entries to avoid duplicates

Do not skip de-duplication. Parallel agents frequently discover the same issues independently.

CLAUDE.md Integration

After the user approves lessons, optionally append to CLAUDE.md:

1. If no CLAUDE.md exists, skip entirely.
2. Find or create a ## Lessons Learned section.
3. Append concise single-line bullets (omit phase dates, focus on actionable guidance).

Selection criteria for CLAUDE.md (be selective — CLAUDE.md has a token budget):

• Actionable every session, not just in this phase context
• Not phase-specific (generic enough to apply to future work)
• Fits as a one-liner (two sentences → split or summarize)
• Not already captured in existing CLAUDE.md entries

Quality Standards

Lessons must be specific, actionable, and reusable. Apply these filters:

Anti-Patterns to reject:

• Lessons that duplicate existing entries in LESSONS.md
• Lessons that reference specific line numbers or ephemeral file locations
• Lessons that are generic truisms rather than discovered knowledge
• Lessons longer than two sentences -- split or summarize

Lesson Quality Examples

Good Lesson -- specific, transferable, actionable

### Pitfalls to Avoid
- bats-core `run` captures exit code but swallows stderr -- use `2>&1` to capture both

Why it works: Names the exact tool and behavior, explains the symptom, and gives the fix.

Good Lesson -- documents a non-obvious decision

### Surprises / Discoveries
- jq `.field // "default"` prevents null propagation in optional config values --
  without the fallback, downstream commands silently receive "null" as a string

Good Lesson -- process improvement example

### Process Improvements
- Run `shellcheck` on test helper files, not just source scripts —
  helper functions have the same quoting bugs as production code

Why it works: Identifies a concrete workflow gap and specifies exactly what to add.

Bad Lesson -- vague platitude

### What Went Well
- Tests are important

Why it fails: Generic truism. Zero discovered knowledge.

Bad Lesson -- too specific, not transferable

### Pitfalls to Avoid
- Fixed a bug on line 47 of parser.py

Why it fails: Line 47 will change. Future readers cannot act on this.

Bad Lesson -- implementation detail, not a lesson

### Process Improvements
- Changed variable name from x to y

Why it fails: A code change, not a reusable insight.

Integration

Called by: commands/ship.md — Step 7 for post-phase lesson capture Pairs with: shipyard:shipyard-verification — validate lesson quality before persisting Leads to: shipyard:documentation — update project docs with any discovered conventions