Quintet Discipline

Three habits that keep multi-agent work honest. Distilled to the load-bearing parts; skip the ceremony on small tasks.

When orchestrating quintet providers, all three matter more, not less: a provider claiming success is not evidence, consensus is not correctness, and a synthesis file existing is not the same as it having content.

---

1. Verification Gate — evidence before claims

Iron Law: no completion claim without fresh verification evidence run this turn. If you didn't run the command now, you can't say it passes.

Before claiming any success:

1. Identify the command that proves the claim.
2. Run it fresh (not cached, not "last time").
3. Read full output — exit code, failure count.
4. Verify the output actually confirms the claim.
5. Then state the claim with the evidence.

Red flags — if you think any of these, stop and run the check: "should work now" · "I'm confident" · "just this once" · "the linter passed" · "the agent said it worked" · "it's a small change".

After a quintet fleet/team run: confirm the output file exists and has real content (wc -l), check timestamps for staleness, and remember exit 0 means the script ran — not that the result is good.

---

2. Intent Contract — lock the goal before building

For any multi-phase task, capture intent up front so output can be validated against it (not against scope drift). Keep it lightweight — a few lines, not a form. Skip entirely for quick single actions.

Capture before starting:

• Job statement — what the user is actually trying to accomplish, plainly.
• Success criteria — "good enough" (minimum) and, if relevant, "exceptional".
• Boundaries — what this should NOT become (over-engineered, out of scope, off-architecture).
• Constraints — stakeholders, existing assets to build on, timeline, technical limits.

For substantial multi-session work, persist it to .claude/session-intent.md so it survives context loss. For a single session, holding it in the task list is enough.

At key decisions, sanity-check against it ("does this still serve the job statement?"). At the end, validate each criterion (met / partial / not) and each boundary (respected / violated) before reporting done.

---

3. Writing Plans — zero context, bite-sized

When asked for a plan, produce a plan — do not jump to implementation, and don't downgrade it to a vague outline because the task feels simple. Rationale and recaps belong in logs/notes, not the plan; a plan is strictly the steps.

Write as if the executor has zero context for the codebase. Document: which files to touch, the actual code, how to test, how to verify.

Task granularity — each task is ONE action (~2–5 min): "write the failing test", "run it to confirm it fails", "implement minimal code", "run to confirm pass", "commit". Not "add the feature".

Per task, give:

• Exact paths — src/validators/email.ts line 23, not "in the utils folder".
• Complete code — the actual snippet, not "add validation logic".
• Exact commands + expected output — npm test path/spec.ts → PASS 3/3.
• TDD order — test before implementation; commit after each green step.

Bottom line: if an engineer with zero context can execute it, it's a plan. Otherwise it's an outline.

---

Persona references

On-demand framing cards in references/personas/ (zero always-on cost). Load one only when the task or the provider prompt you're about to write falls in its domain:

• When auditing auth, secrets, input handling, or attack surface → read references/personas/security-auditor.md.
• When debugging deploys, networking, CI, or infra incidents → read references/personas/devops-troubleshooter.md.
• When isolating a regression or chasing a root cause → read references/personas/debugger.md.

Do NOT load them for general implementation or planning work — they're reference prompts to shape one analysis, not registered agents and not always-on context.