AI Governors — HITL product UX

Governors are product patterns that keep the owner in control while the agent acts. Mechanism skills answer "is the gate sound?". This skill answers "which oversight shape should the surface expose?".

Pattern catalog: references/patterns.md. Adapted from tommyjepsen/awesome-ux-skills ai-governors.

When to use

• Write-approval queues, staged proposals, action confirm, spend UI, agent activity / runs, memory capture, voice speaker gates.
• Phrases like: confirm before act, undo, show cost, action plan, cite sources, draft then commit, pause the agent.
• Not for: low-level gate wiring audits, untrusted prompt framing, or workflow Send-and-Wait plumbing in isolation.

Seam → pattern map

Seam	Prefer	Hard rules
Staged writes / approvals	Verification + Memory + Footprints	Queue must drain; bind approval to exact bytes
Proposal accept/reject	Verification + Sample + Cost + Draft	Never auto-apply on model confidence alone
Action / code confirm	Action Plan + Verification + Cost	Confirm ahead of irreversible fire
Cloud / research spend	Cost Estimates (+ Draft for bakeoffs)	Unknown cost → visible, not "free"
Runs / agent activity	Shared Vision + Stream + Controls	Stop/cancel only if real control exists
Memory / fact capture	Memory (+ Verification when sensitive)	Secret scan before store
Research / web / vision	Citations + References	Untrusted framing; honest observer mode
Voice / mic gates	Controls + Verification	Partial API must not crash UI

Calibration (friction vs speed)

Match friction to irreversibility × blast radius:

1. High (money, outbound message, memory write, destructive tool) → Action Plan or Verification; no silent defaults that grant the act.
2. Medium (long research, multi-step run) → Shared Vision + Controls + Cost; Verification on the commit step.
3. Low (search, draft reply) → skip Verification; optional Citations.

Prompt fatigue is a defect: confirming every chat turn trains click-through. Prefer one strong gate on the irreversible step.

Design checklist

• Name the irreversible effect in the confirm copy.
• Approval bound to exact bytes / run id / proposal id — not vague intent.
• Cost or side-effect visible before commit when spend is non-trivial.
• Undo or compensating action stated when reversible; say so when not.
• Failure path does not skip the gate.
• Absent evidence ≠ consent.

Anti-patterns

• Manifest that promises a capability the confirm gate still blocks.
• Staging then auto-applying "because the model was confident".
• Cost UI that meters one lane while another is silently free.
• A Stop control that only navigates away.
• Partial state payloads that crash the oversight UI.

Output format

1. Situation — blast radius in one sentence.
2. Patterns — 1–3 from the catalog.
3. Surface — concrete UI / flag / API.
4. Friction — high / medium / low and why.

Cross-references

• ai-trust-builders — disclosure beside oversight
• ai-wayfinders — first-run discoverability (does not replace HITL)