| name | zero-defect |
| description | Session-scoped precision mode — read-before-write, verify-before-claim, never-hallucinate-APIs. Use when stakes are high (security, payment, migration). Triggers: 'zero defect', 'precision mode', 'high stakes', 'be careful'. |
| context | shared |
| effort | medium |
/zero-defect — Precision Mode
Inspired by Maestro /zero-defect. Switches the agent into a stricter operating mode for the rest of the session.
When to use
- Touching auth, payments, billing, crypto, secrets, session handling
- Database migrations or schema rebuilds
- Bootstrap / packaging changes that ship to all users
- Any task tagged with
complexity=complex and defect_of non-empty (defect-fix)
The 8 Rules
For the rest of this session:
- Read before write — every Edit must be preceded by Read of the same file in the same turn (or a verified prior Read).
- Verify before claim — never write "tests pass" / "feature works" without running the test or operation in this session.
- Don't hallucinate APIs — if uncertain, run
mcp__codebase-rag__search_code (RAG) for the symbol; fall back to Grep only if RAG is empty or stale. Then read upstream docs.
- Re-derive don't recall — for tricky logic, re-derive from current code state; don't trust memory of a previous read.
- Smaller edits — prefer many small Edits with verification between, over one large rewrite.
- Atomic commits — group changes by concern; never bundle a refactor with a feature.
- Single responsibility per task — split if scope creeps.
- Pre-commit gate — run
/test and /review (TAUSIK's 6-agent pipeline) before saying "done".
Activation
When user invokes /zero-defect:
- Acknowledge mode is active and list the 8 rules.
- For the rest of the session, prepend each substantive response with:
[ZERO-DEFECT].
- Refuse to mark
task done --ac-verified without recent test evidence in notes.
Gotchas
- Slows velocity by ~2-3×. Use only when warranted, not as default.
- Doesn't replace QG-2 — runs IN ADDITION to gates.
- Cannot enforce rule 1 (Read-before-Write) at the framework level — relies on agent discipline.