Skills Quality assurance
Human-in-the-loop design
Every approval gate adds latency, labour cost, and a new failure mode: agents approve without reading because the queue is backed up. The design question is not "should humans be involved?" but where involvement changes outcomes enough to pay for itself.
npx rulebase-skills install cx-human-in-the-loop-designWhen to use it
Reach for this when someone says any of these — they are the phrases the skill itself triggers on:
- “human in the loop design”
- “where should agents approve AI actions”
- “approval workflow for the bot”
- “AI auto-send vs review”
- “rubber stamping risk”
How it works
The method, in the order the skill runs it. The full procedure — tables, worked examples and the edge cases — is in the skill itself.
Classify actions by reversibility and blast radius
The gate belongs on the tier boundary, not on every assist feature. A copilot draft is tier 2 — the send click is the gate. A bot that posts refunds is tier 4 — model output is never sufficient.
Where to place the gate
Options for GATE placement.
Latency and cost trade-off
Estimate for each gate design: Seconds added per conversation × volume × agent cost.
Rubber-stamping failure mode
Rubber-stamping happens when: SLA pressure exceeds review capacity, Incentives reward speed over quality.
What humans should decide vs what machines should rank
Use humans for accountability and ambiguity; use models for drafting and triage — not the reverse.
Document the policy
For each channel and action type, record: Auto-allowed / assist-only / confirm-required / forbidden, Confidence or retrieval thresholds if used, Owner for exceptions.
Related skills
Free and open source, and vendor-neutral — it reads the conversations from whichever helpdesk you already run. Browse all 149 skills · connect your helpdesk over MCP · source on GitHub
