> workflow / human-review
Human Review
Approve the implementation plan (and risk posture) before any AI coding — a plan-approval HITL gate, not a pull-request review.
What should I do next?
Complete the steps and evidence below, clear the human gates, then continue to AI Coding.
When to use
- Implementation plan file is ready and claims DoR for coding
- Risk tier requires accountable human sign-off before agents write code
- Architecture boundaries or debt acceptance need explicit human yes/no
- You must record who approved what before pre-code hooks unlock
Step-by-step
- Review the plan against the brief metrics — reject work that does not move them.
- Check every slice for ADR alignment, tests, rollback, and owners.
- Challenge AI-generated optimism: missing migrations, permissions, and on-call impact.
- Confirm MCP/tool allowlist for the upcoming coding stage is still correct.
- Record approval (or change requests) in the plan PR / change log — not in chat.
- Optionally capture decisions with knowledge-capture.prompt for later retros.
- Unlock coding only after approval; leave pre-code.hook.sh blocking until then.
Decisions to make
- Approve as-is vs request plan revisions
- Which slices may use higher AI autonomy vs pair-required
- Whether elevated risk needs security co-approval before coding
- Waiver (time-boxed) vs hard block for incomplete evidence
Evidence to collect
- Signed plan approval with named approver and timestamp
- Resolved change requests or explicit waivers with expiry
- pre-code gate configuration reflecting approval state
- Updated risk notes if plan changed architecture assumptions
Role-specific next actions
- PM: Sign that slices still deliver the brief; cut scope rather than rubber-stamp.
- Architect: Veto slices that violate ADR boundaries or invent integrations.
- Engineer: Confirm the plan is executable; refuse vague "AI will figure it out" steps.
- Engineering Manager: Ensure segregation of duties: planner ≠ sole approver on high-risk work.
Inputs
- Implementation plan file and task graph
- Risk tier and ADR set
- Reviewer roster / approvers
Outputs
- Recorded plan approval (or rejected with required changes)
- Autonomy level per slice
- Unlocked pre-code gate for approved slices only
Required prompts
- Implementation plan
Re-run or diff against the living plan under review — this stage approves plans, not PRs.
- Knowledge capture
Optional: capture why the plan was approved or cut for later improvement.
Required skills
- Planning
Primary skill — evaluate plan completeness, not PR diffs.
Required MCP servers
None required at this stage — keep tools off by default.
Tooling & handbook
- Human in the loop
Defines plan-approval HITL gates and accountability.
- How to set up planning
Plan artifact and gate mechanics reviewers check.
- pre-code hook
Must stay blocking until plan approval is recorded.
- AIDLC roles & evolving delivery
Who may approve vs who may execute.
Recommended agents & patterns
- Planner pattern
Primary pattern — plan quality and decision completeness.
- Planner agent
May summarize plan risks; humans approve.
Human approval gates
- Named human approves the implementation plan before coding
- High-risk tiers require architect (and security when data/tools expand) co-sign
- No production-impacting AI coding without this gate
Automation opportunities
- Plan PR template requiring ADR links, tests, rollback
- pre-code.hook.sh reads approval status from plan metadata
- Notify approvers via existing PM tooling after plan PR opens
Enterprise best practices
- Segregation of duties on high-risk plans
- Retain approval evidence for audits
- Time-box waivers; expired waivers re-block coding
Common mistakes
- Treating this stage as PR/code review (that is the code-review stage)
- Rubber-stamping AI-written plans without checking rollback and tests
- Using reviewer-pattern / code-review.skill as the primary gate here