Project learnings

Project Learnings

Capture project-specific patterns and anti-patterns into the project's CLAUDE.md. This creates a self-improving feedback loop where discoveries from debugging, development, and review make future Claude sessions smarter.

CRITICAL: Only project-specific knowledge qualifies. Generic programming advice does not belong in CLAUDE.md.

---

Step 1: Evaluate Discovery

Determine if the finding qualifies as project-specific. The finding must pass at least ONE of these criteria:

Criteria	Example That Qualifies	Example That Doesn't
Would a developer unfamiliar with this project likely hit this issue?	"The processOrder() function expects amounts in cents, not dollars"	"Always validate function inputs"
Is this pattern specific to this codebase's architecture, APIs, or conventions?	"The UserProfile type has an optional metadata field that is always present at runtime"	"Use TypeScript strict mode"
Is it something Claude's training data wouldn't cover?	"Never call db.query() without the timeout option — the default is infinite"	"Use async/await instead of callbacks"

If NO to all criteria → STOP. Do not add generic programming knowledge to CLAUDE.md. Return to the calling skill and report that no project-specific learning was found.

If YES to any → proceed to Step 2.

---

Step 2: Read Existing CLAUDE.md

1. Find the project's CLAUDE.md:

   • Check the repository root first
   • If not found, check if there's a project-level .claude/ directory

2. Parse existing content:

   • Understand the existing structure, headings, and conventions
   • Look for sections where this learning would fit (e.g., "Known Gotchas", "Bug Patterns", "Conventions", "Known Challenges")
   • Check for duplicate or similar entries already present

3. If a similar entry already exists → STOP. Report to the calling skill that this knowledge is already captured. Do not create duplicates.

4. Identify placement:

   • If an appropriate section exists, plan to add the entry there
   • If no appropriate section exists, plan to propose a new section (e.g., ## Known Gotchas or ## Project-Specific Patterns)
   • New sections should be placed after the main documentation sections but before appendices or settings

---

Step 3: Format the Learning

Write a concise, actionable instruction following these rules:

Format:

• Use imperative form: "Always validate X before calling Y"
• Include the WHY: "...because the API returns dates as strings, not Date objects"
• Keep it to 1-3 lines
• Follow the existing CLAUDE.md style and conventions

Templates:

For bug patterns:

- **[Area/Component]**: [What to do/avoid] — [why, with specific details]

For API gotchas:

- `functionName()` in `path/to/file`: [What's surprising about it] — [consequence if ignored]

For architectural constraints:

- [Constraint description] — [why it exists and what breaks if violated]

Examples of well-formatted learnings:

• Order processing: Always multiply amounts by 100 before passing to processOrder() — it expects cents, not dollars
• db.query() in src/database.ts: Always pass the timeout option — the default is infinite and has caused production hangs (30 second timeout recommended)
• Never import from internal/ directories in src/api/ — the build system treats these as separate compilation units and circular dependencies will silently break HMR

---

Step 4: Confirm with User

Present the proposed addition via AskUserQuestion:

Show:

1. The exact text to be added
2. Where it will be placed in CLAUDE.md (section name, after which line/entry)
3. Why this qualifies as project-specific

Options:

• "Add this" — Write the entry as proposed
• "Edit before adding" — User provides modified text, then write that instead
• "Skip" — Do not add anything, return to calling skill

---

Step 5: Write Update

If the user confirmed (or provided edited text):

1. Use the Edit tool to add the entry to CLAUDE.md at the identified location
2. If a new section was needed, create the section heading first
3. Verify the edit was applied correctly by reading the modified area
4. Report success to the calling skill with a summary of what was added

If the user chose "Skip":

• Report to the calling skill that the learning was declined
• Do not modify any files