| name | planning |
| description | Produce implementation plans that are decision-complete and execution-ready. Use when the task needs decomposition, sequencing, acceptance criteria, explicit assumptions, risk handling, or a concrete plan before coding or operational work starts. |
Planning
Overview
Turn ambiguous work into a plan another agent can execute without reopening core decisions. Resolve scope, sequence, constraints, and definition of done before treating the plan as complete.
Output Standard
Produce plans that answer all of these points:
- Goal: what will be true when the task is done.
- Scope: what is included and what is explicitly excluded.
- Assumptions: facts treated as true until disproved.
- Dependencies: tools, services, approvals, inputs, or files required.
- Sequence: the ordered work with critical-path awareness.
- Acceptance criteria: observable checks for completion.
- Risks: likely failure modes and how to handle them.
- Open questions: only if they block a safe decision.
Do not leave the plan at a vague "investigate" level unless the task itself is discovery work.
Workflow
1. Frame The Request
- Restate the intended outcome in concrete terms.
- Identify the artifact to produce: code change, migration, document, deployment, analysis, or decision.
- Distinguish hard constraints from preferences.
- Call out missing inputs that materially affect the path.
2. Define Boundaries
- State what the plan will cover.
- State what it will not cover.
- Convert broad asks into a smallest useful deliverable if full scope is unclear.
- Prefer one coherent scope over a mixed list of loosely related tasks.
3. Decompose The Work
- Break the task into units that have a clear completion signal.
- Separate setup, implementation, validation, rollout, and cleanup when relevant.
- Identify which steps can run in parallel and which are blocking.
- Keep decomposition shallow unless the task is large or high risk.
Good step wording:
- "Inspect the current auth flow and list the request/response contracts."
- "Add the database column and backfill existing rows."
- "Update the API handler and client call sites to use the new field."
- "Run focused tests for migration and regression coverage."
Weak step wording:
- "Handle backend"
- "Do frontend stuff"
- "Test everything"
4. Sequence For Execution
- Order steps so each one unlocks the next with minimal rework.
- Put irreversible or high-blast-radius work after validation where possible.
- Move expensive steps later if cheaper checks can fail first.
- Call out prerequisites before dependent implementation steps.
5. Make Decisions Explicit
- Choose defaults when the evidence is good enough.
- Record the decision and brief rationale.
- Surface tradeoffs only when they affect implementation or risk.
- Ask a question only when the answer changes the plan materially and a safe assumption is not available.
6. Define Acceptance Criteria
- Write criteria that are externally verifiable.
- Prefer outcomes over activity.
- Tie each criterion to behavior, data state, interface contract, or delivery artifact.
- Include validation commands or checks when known.
Examples:
- "The endpoint returns
409 for duplicate slugs and preserves the existing response schema for other cases."
- "The migration completes on existing data without null violations."
- "The page renders correctly on mobile and desktop with no console errors."
7. Address Risks And Unknowns
- List the few risks that are most likely or most costly.
- Pair each risk with a mitigation, checkpoint, or fallback.
- Mark unknowns as blocking only when proceeding would be unsafe or wasteful.
- If discovery is required, define the exact question to answer and the artifact expected from that discovery.
Decision Rules
- Prefer a smaller plan that can be executed now over a larger plan with unresolved core choices.
- Prefer concrete nouns and verbs over category labels.
- Prefer acceptance criteria that can fail clearly.
- Prefer assumptions plus caveats over unnecessary back-and-forth.
- Revise the plan if new facts invalidate sequencing or scope.
Plan Shapes
Use the lightest structure that still removes ambiguity.
Small Task
Use a short plan with:
- outcome
- 3-5 ordered steps
- acceptance criteria
Medium Task
Use:
- outcome and scope
- assumptions and dependencies
- ordered implementation steps
- validation steps
- risks
Large Or High-Risk Task
Use:
- outcome, scope, and non-goals
- architecture or rollout decisions
- phased sequence with dependencies
- validation matrix
- rollback or fallback approach
- unresolved questions, if any
Completion Bar
A plan is decision-complete when a competent implementer can start work without asking:
- what to do first
- what success looks like
- what assumptions are in force
- which risks need active handling
- whether major product or technical decisions are still pending
If those answers are missing, the plan is not finished.