AI-Human Collaboration Flow
Standard Collaboration Pattern
The canonical workflow is: AI Draft -> Human Review -> AI Revise
mermaid
flowchart LR
A[AI Drafts] --> B[Human Reviews]
B -->|Approved| C[AI Finalizes]
B -->|Feedback| D[AI Revises]
D --> B
C --> E[Outcome Recorded]
Decision Recording
| Decision Type | Record Location | Promotion Target |
|---|---|---|
| Task-specific decisions | workdocs/ (plan.md, context.md) | Archive when complete |
| Stable knowledge | workdocs/ initially | Promote into the relevant skill package under /.system/skills/ssot/** |
| Repeated patterns | workdocs/ observations | Abilities or workflows |
| Policy decisions | Review comments | Update AGENTS.md if applicable |
Review Feedback Format
Humans provide review feedback using inline comments or structured blocks:
markdown
<!-- REVIEW: [reviewer_name] [date] Status: approved | needs_revision | blocked Feedback: Specific feedback text here -->
AI must:
- •Parse and address all needs_revision feedback.
- •Not proceed past blocked status without human resolution.
- •Update workdocs with review outcomes.
Human Approval Gates
Certain actions require explicit human approval:
| Action Type | Approval Required | Record Method |
|---|---|---|
| Production deployment | Yes | workdocs/tasks.md checkpoint |
| Security-sensitive changes | Yes | workdocs/tasks.md checkpoint |
| Cross-module breaking changes | Yes | Integration scenario workdocs |
| Credential/secret access | Yes | DevOps scenario workdocs |
AI must stop and request approval when:
- •A hook returns requires_human.
- •An action is in the "Forbidden without approval" list in applicable AGENTS.md.
- •Uncertainty exists about safety implications.