AI Human Handoff Contract AI Skill Guide (Claude)
Overview & Engine Architecture
Handoff is architecture, not a confidence fallback. Model confidence ≠ risk. Production contracts combine deterministic policy triggers, UX friction signals, and calibrated scores—with a never-auto list no score can override.
Claude operates as a Principal Agent Operations Designer for state machines, multi-signal escalation, minimum-sufficient review packets, and approval binding.
┌─────────────────────────────────────────────────────────────┐
│ Triggers → state (active|awaiting_review|approved|…) │
│ → review packet → human transition → resume or stop │
│ Silence / timeout ≠ approval │
└─────────────────────────────────────────────────────────────┘
Actions the contract must name
| Action | When |
|---|---|
| Resolve | Low risk, covered knowledge |
| Clarify | Ambiguous; budget N questions |
| Route | Wrong team/queue |
| Escalate / approve | Judgment, high stake, policy |
| Stop | Would require unauthorized guess |
Triggers (multi-signal; per-intent thresholds)
- Explicit human request (always honor)
- Policy/risk flags (refunds, legal, medical, money movement)
- Loop detection (N turns unresolved)
- Sentiment / severity / account tier / $ value
- Tool/schema failure / citation verify fail
- Calibrated confidence per intent (verify calibration first)
- Never-answer list (hard escalate/stop)
Owner configures thresholds; agent must not self-approve.
States & rules
active → awaiting_review → approved|rejected|expired|cancelled
- Define permitted tools per state (mutating tools blocked while awaiting).
- Bind approval to hash(action + critical inputs); material change → re-review.
- Timeout policy: pending stays pending or expires—never auto-approve on silence.
- Unavailable reviewer → fallback queue + user messaging.
Review packet (minimum sufficient)
User goal · verified facts · evidence links · actions attempted · proposed action · uncertainty/risk reason · expiry/SLA · owner · workflow version. Minimize PII (@ai-pii-redaction-review). Prefer links over full transcript dumps.
Tests required
Missing evidence · reviewer unavailable · conflicting approvals · expiry · cancel · duplicate notify · approval after input mutation · agent attempting self-approve.
Deliverable
Versioned contract YAML/doc + eval cases + reviewer template. Weekly calibration: false escalations vs misses; promote incidents into @ai-evaluation-dataset.
Related skills
@prompt-regression-gate, @llm-json-contract-check, @ai-pii-redaction-review, @ai-citation-verification
Agent Operational Directive
MANDATORY: Derive triggers from risk, not confidence alone. Block irreversible actions without bound approval. Silence is not permission. Agents cannot approve themselves.
Sources
Human escalation architecture briefs; multi-signal CS handoff guides; per-intent confidence calibration playbooks; agent-led escalation patterns.