> workflow / implementation-planning

Implementation Planning

Break approved architecture into verifiable slices with tests, rollback, and evidence hooks — before any feature coding begins.

What should I do next?

Complete the steps and evidence below, clear the human gates, then continue to Human Review.

When to use

  • ADRs and boundary freeze exist
  • Work must be sliced for parallel humans/agents without scope thrash
  • You need a plan file that pre-code hooks can enforce
  • Test strategy and rollback must be explicit before coding

Step-by-step

  1. Derive slices from ADRs and contracts — each slice shippable with evidence.
  2. For each slice: files touched, tests, evals, rollback, and Definition of Done.
  3. Order the task graph by dependency and risk; put spike slices first when unknowns remain.
  4. Call out technical debt you will deliberately take and debt you refuse.
  5. Run implementation-plan.prompt; produce a plan file in-repo (not a chat transcript).
  6. Attach test strategy outline and eval gates for AI-touched paths.
  7. Stop at plan completion — do not code until human-review approves the plan.

Decisions to make

  • Slice size: hours vs days — prefer slices that fit one PR
  • What is spike vs committed delivery
  • Which debt is accepted with an expiry date
  • Which slices require pair programming vs solo AI-assisted work

Evidence to collect

  • Committed implementation plan file with task graph
  • Test/eval strategy linked per high-risk slice
  • Rollback steps for each production-touching slice
  • pre-planning.hook.sh pass on the plan artifact

Role-specific next actions

  • PM: Validate slices still map to brief metrics; cut or reorder ruthlessly.
  • Architect: Reject slices that violate ADR boundaries or invent new integrations.
  • Engineer: Estimate and challenge missing tests, migrations, and observability work.
  • QA: Own test strategy outline and risk-based coverage targets.

Inputs

  • Approved ADRs and contracts
  • Backlog / capacity
  • Risk tier and DoR

Outputs

  • Implementation plan file
  • Task graph with owners
  • Test and eval strategy outline

Required prompts

Required skills

  • Planning

    Primary skill — decision-complete execution contracts.

  • Testing

    Ensures each slice has verifiable proof.

Required MCP servers

  • Jira MCP

    Optional: sync plan slices to tickets after plan approval.

  • Linear MCP

    Optional alternative PM sync — never invents scope.

Tooling & handbook

Recommended agents & patterns

Human approval gates

  • Plan file exists in-repo and passes pre-planning checks
  • No coding until the human-review (plan approval) stage completes

Automation opportunities

  • pre-planning.hook.sh on plan PRs
  • Ticket creation from approved slices only
  • Link each slice to ADR IDs in the plan file

Enterprise best practices

  • Plans reference control IDs where regulated workloads apply
  • Capacity and dependency risks explicit — no hidden platform work
  • Retain plan versions for audit when AI executes slices

Common mistakes

  • Monolithic plans that cannot be reviewed or rolled back
  • Missing rollback and observability steps
  • Coding from a chat plan that never landed in git