stas2-git Human Approval Gate Designer
Use this skill to make approval meaningful instead of a vague "Are you sure?"
Workflow
- Identify actions that need approval: external write, irreversible change, sensitive data, production action, payment, or ambiguous intent.
- Define what the reviewer must see before deciding.
- Define allowed decisions: approve, deny, edit, request clarification, escalate.
- Define what happens after each decision.
- Add audit fields and replay/session-resume requirements.
- Set auto-approval only for low-risk, bounded, reversible actions.
Approval Payload
Include:
- action summary,
- exact target,
- proposed payload or diff,
- reason the agent thinks it is needed,
- risk level,
- data sources used,
- policy checks passed/failed,
- rollback or undo path,
- expiration/time limit.
References
- Read
references/gate-patterns.mdfor payload templates, risk tiers, and anti-patterns.
Evaluation Prompts
- Positive: "Design an approval flow before this agent sends customer emails." Expected: payload, decisions, audit, send boundary.
- Positive edge: "Some deploys should auto-approve." Expected: tiered gates and rollback criteria.
- Negative: "Make a UI button say Approve." Expected: no skill unless approval policy/flow is being designed.