design-human-ai-workflow
Installation
SKILL.md
Design Human AI Workflow
Contract
Design the joint system, not an AI feature in isolation. Autonomy must match uncertainty, reversibility, user expertise, and consequence. A human approval step is not a control when the reviewer lacks context, time, or a meaningful alternative.
Workflow
- Map the current job, actors, information, tools, decisions, handoffs, exceptions, incentives, and failure recovery.
- Identify where AI could retrieve, transform, recommend, generate, decide, or act. Compare it with simpler automation and unchanged human work.
- Assign responsibility for each step: human only, AI proposes, AI acts with confirmation, AI acts within limits, or AI acts with retrospective review.
- Define context sources, freshness, permissions, provenance, uncertainty, and what the system must never infer or retain.
- Design previews, explanations, confidence or evidence cues, editing, approval, cancellation, undo, escalation, and safe defaults appropriate to the risk.
- Handle ambiguity, tool failure, partial completion, conflicting instructions, stale state, unavailable reviewers, and repeated failure without silent continuation.
- Define learning and feedback loops without making users responsible for endless correction or turning unreviewed feedback into truth.
- Measure task success, effort, time, quality, override, recovery, trust calibration, distributional effects, and automation-induced errors.