---
name: stop-conditions
description: "Defines observable completion, budget, approval, failure, and safety terminal conditions for an agent loop. Use before autonomous, repeated, expensive, or production-adjacent work."
---

# Define stop conditions first

Apply this skill during **Architecture** work. It supports Any stack.

Make done, blocked, unsafe, and not-worth-continuing observable before the loop starts.

## Instructions

Define terminal conditions before starting this loop. Describe done as an observable outcome plus the evidence required to claim it. Set explicit caps for time, cost, turns, retries, and scope. List actions that must stop for approval: destructive changes, secret access, external communication, production mutation, legal or financial commitment, or any expansion of authority. Define a repeated-failure threshold and what materially different approach must follow it. Define when weak evidence means narrow or stop rather than continue producing. At termination, require a report of evidence collected, gaps remaining, changes made, and the safest next action.

## Detect the problem

- Keep going until it feels finished
- The same failed approach retried with different wording
- A budget exists but has no behavior when it is reached

## Hold the gate

Do not declare this phase complete until every item passes:

- [ ] Done is an observable outcome with required evidence, not effort spent
- [ ] Time, cost, turn, and scope budgets have explicit caps
- [ ] Unsafe, destructive, external, or authority-expanding actions require a stop or approval
- [ ] Repeated-failure and no-progress thresholds define when to change approach
- [ ] The final state requires a concise evidence, gap, and next-action report

## Verify the result

Present five scenarios—done, budget exhausted, repeated failure, approval required, and weak evidence—and confirm each produces one unambiguous terminal action.