English | Русский

/zero-defect — Precision Mode

Vendor skill — not shipped with bootstrap by default. /zero-defect lives in the public tausik-skills repo as part of the quality-pro bundle. Install it with .tausik/tausik skill install zero-defect (single skill) or .tausik/tausik skill bundle install quality-pro (zero-defect + audit + security + optimize + ultra). See Skill Bundles.

/zero-defect is a session-scoped operating mode that tightens agent discipline for high-stakes work. It does not promise bug-free output — it lowers the rate at which careless mistakes slip through.

The skill is inspired by Maestro's /zero-defect, adapted to TAUSIK's QG-0 / QG-2 model.

When to Use

Trigger it explicitly for tasks where a single mistake is expensive:

• Security surface — auth flows, JWT/session handling, password reset, RBAC
• Money — payment intent creation, refund/void, ledger writes, subscription state
• Migrations — schema rebuilds, data backfills, irreversible cleanups
• Bootstrap / packaging — anything that ships to every user
• Defect fixes on complex tasks — complexity=complex and defect_of non-empty

Activate by saying zero-defect, precision mode, high stakes, or be careful. The skill recognises these triggers and toggles on.

The 8 Rules

For the rest of the session, the agent commits to:

1. Read before write — every Edit is preceded by Read of the same file in the same turn (or a verified prior Read).
2. Verify before claim — no "tests pass" / "feature works" claim without running the test or operation in this session.
3. Don't hallucinate APIs — when uncertain, grep the codebase or read upstream docs before invoking.
4. Re-derive, don't recall — for tricky logic, re-derive from current code; don't trust memory of a previous read.
5. Smaller edits — many small Edits with verification between, not one large rewrite.
6. Atomic commits — group by concern; never bundle a refactor with a feature.
7. Single responsibility per task — split if scope creeps.
8. Pre-commit gate — run /test and /review (TAUSIK's parallel review pipeline) before saying "done".

Activation Behaviour

When /zero-defect is invoked:

1. The agent acknowledges precision mode and re-states the 8 rules.
2. For the rest of the session each substantive response is prefixed with [ZERO-DEFECT].
3. The agent refuses to call task done --ac-verified without recent test/operation evidence in notes.

Cost — and Why That's the Point

• Velocity drops by ~2–3× compared to standard mode.
• Tool-call budgets should be increased proportionally.
• The gain is fewer escapees: defects that would have shipped, the user noticing only after deploy.

Negative — What /zero-defect Does Not Promise

• It does not promise bug-free output. It promises a tighter loop, not a perfect one.
• It does not replace QG-2. Quality gates still run; /zero-defect runs in addition.
• It cannot be enforced at the framework level for rule 1 (Read-before-Write). Agent discipline carries it.
• Default mode for casual tasks is faster. Don't switch on /zero-defect unconditionally — it has a cost.

Combine With

• /review — runs the 5-agent SENAR review pipeline; /zero-defect rule 8 invokes this
• /test — explicit test execution; rule 2 requires evidence
• QG-2 scoped pytest gate — task done already runs scoped tests; /zero-defect requires running them before calling task done

What's Next

• Skills — full skill catalogue
• Workflow — when precision mode fits the day
• SENAR Compliance Matrix — how this composes with QG-0 / QG-2