PROMPTS.md

# PROMPTS.md
> Prompt library. What worked and what did not.
> Add every prompt pattern that produced good output. Reuse relentlessly.

---

## High-signal prompt patterns

### Starting a session
    Read HANDOFF.md (last entry only). Read 2-context/CONSTITUTION.md.
    Confirm your understanding of the current task before writing any code.
    Current task: [TASK DESCRIPTION]

### Specifying a feature
    Read 3-workflow/blueprints/[FEATURE].md.
    Implement this feature exactly as specified.
    Do not add anything not in the spec.
    Ask me before deviating from any acceptance criterion.

### Security review
    Review this code as a security engineer, not a developer.
    Check against: OWASP Top 10, the checklist in 4-security/SECURITY.md.
    List every issue you find, ranked by severity.
    Do not fix anything yet - list first.

### Debugging
    Do not fix this yet.
    First: explain what you think is causing this bug.
    Second: list the files involved.
    Third: describe the minimal change that fixes it without touching anything else.
    Wait for my confirmation before writing any code.

### Preventing scope creep
    This task is ONLY about [X].
    Do not refactor anything outside the direct scope.
    Do not improve things you notice.
    Do not add error handling beyond what the task requires.

### Deep thinking
    Think hard about this before responding.
    What are the non-obvious failure modes?
    What would a senior engineer push back on here?

### Updating intent
    You just made a non-obvious decision.
    Add an entry to 2-context/INTENT.md now, before we continue.
    Format: date, component name, decision, why, what not to change.

### Closing a session
    Write a handoff entry for this session.
    Use the template at the top of 3-workflow/HANDOFF.md.
    Then update 3-workflow/PLAN.md with what was completed and what is next.

---

## Known anti-patterns (prompts that produce bad output)

These fail because they give the AI too much latitude.

- "Improve this code" - AI refactors things that were intentional
- "Clean this up" - AI deletes comments and renames variables, loses intent
- "Make this better" - undefined success criteria, AI guesses
- "Fix the bug" without specifying scope - AI rewrites surrounding code
- "Add error handling" without specifying which errors - AI adds catch-all handlers
- "Continue where we left off" - AI has no memory, will fabricate context

---

## What to add here

When a prompt produces unexpectedly good output, paste it here with a one-line label.
When a prompt produces bad output twice, add it to the anti-patterns section.