The Loop — Wall Card

Scannable in 20 seconds. The cycle every piece of work runs, from a one-line fix to a whole feature. Print it, pin it. The full method is in the standard; the build-loop deep-dive is the long version of this card.

Paired source: docs/cheatsheet.md · part of MCKRUZ/intent-driven-development v1.0

Intent → Delegate → Discern

  INTENT             DELEGATE            DISCERN
  ┌────────┐         ┌────────┐          ┌─────────┐
  │ decide │────────▶│  plan  │─────────▶│  check  │
  │ write  │         │ build  │          │  merge  │
  └────────┘         └────────┘          └─────────┘
   you own this       agent builds,       gates + a human
   (the spec)         you set the box     own the verdict

Beat	Who owns it	The two moves
Intent	Human (Pod Lead owns quality; PO owns the call)	decide what you want, then write it as a spec you can test
Delegate	Agent builds; Orchestrator sets the bounds	agent plans (you approve before code), then builds and self-checks
Discern	Gates check; a non-author human approves	check against the spec, then merge without breaking `main`

Discern just means working out whether the work is actually good before anyone trusts it.

The checking ladder (quick → thorough)

A change climbs as far as its risk demands. Each rung is harder to skip than the last.

A rule in CLAUDE.md about what "done" means.
A check that re-runs every turn to confirm it.
A blocking Stop hook — won't let the agent finish while tests are red. Not optional, not persuadable.
A separate grader that reads the spec and grades the diff. The author never grades its own work.
A human gate for HIGH risk — a security-reviewer pass first, then a named sign-off in the PR.

Mechanical gates (build, tests, lint, 80% coverage) block. The grader runs but advises. The human Checker decides. Machines gate the mechanical; humans own the judgment.

Risk sets the leash

Tier	Examples	Review
HIGH	auth, payments, PII, migrations, public API / pipeline / IaC, prompt-model-tool changes	full ladder + security pass + named human sign-off
MEDIUM	new business logic, external integrations, shared services	grader + human Checker
LOW	UI in existing patterns, copy, internal tooling, additive CRUD	lighter review; grader + mechanical gates still run

Read a typo fix lightly. Read an auth change hard. Challenges to a tier escalate up, never down.

Do

Write the goal as a concrete, testable spec before any code exists.
Set the box the agent works in: scope, the one pattern to reuse, one agent or several.
Run every change through all three beats — even the one-liners.
Match review depth to the risk tier.
Let a non-author grade the work; keep the spec file in the PR diff.

Don't

Don't ship what nobody checked. That's vibe coding to prod.
Don't let the author approve its own work.
Don't split tightly coupled code across many agents — they clobber each other.
Don't open more streams than the pod can review (watch median review wait).
Don't count PRs, lines, or velocity. They go up while delivery stays flat.

No checking is just vibe coding. The loop closes it. → Next: the build-loop deep-dive.