Imported from nmarcofernandess/superflow (plugins/superflow/skills/plan/SKILL.md). Install upstream with npx skills add nmarcofernandess/superflow --skill plan. Copyright stays with the author.

Plan

Plan converts the source of truth into ordered executable work. It should be smaller than the PRD and more executable than the blueprint. Build decides the architecture; Plan creates the task catalog.

Procedure

1. Read ../../assets/references/execution-contract.md.
2. Read ../../assets/references/tdd-contract.md (I1 — plan pre-compiles proof).
3. Read ../../assets/references/status-schema.md.
4. Read local PRD.md first. Then read SPEC.md (or legacy technical_blueprint.md), analysis.md / ANALYSIS-*.md, or repo-native equivalents when present.
5. Read ../../assets/templates/implementation_plan.json before writing a local plan.
6. Write implementation_plan.json as the executable task source. Optionally write implementation_plan.md only as a human-readable summary.
7. Update status.json: phases.plan = "complete", artifacts.plan = "implementation_plan.json", and task_source.path = "implementation_plan.json".
8. Create the human-facing task board: copy ../../assets/task-board/board.html into the package and write board-data.js from the plan tasks (schema in ../../assets/task-board/board-data.example.js). The board is a projection of the plan for the human; it never replaces implementation_plan.json. Skip it only for trivial direct execution with one or two obvious steps.

Required Plan

• Preconditions.
• Ordered tasks.
• File targets.
• Change per task.
• TDD pre-compile per code task (I1): behavior, tdd.red.command, tdd.red.expected_failure, tdd.green.command. Prefer real test file and test name when known.
• Validation per task (verification may mirror tdd.green).
• Done criteria mapped back to PRD acceptance criteria (acceptance_criteria on each task).
• Owner classification per task: main_agent, explorer, worker, or reviewer. A reviewer task runs the review skill and writes review_log.json — the role is a phase now, not a label (../../assets/references/review-contract.md).
• status: "pending" for every task at plan creation. Execution progress goes to implementation_log.json, not into the plan.

TDD rules (summary)

• Code tasks: tdd.required: true by default. Iron law is enforced at Execute.
• Docs/chore-only tasks: tdd.required: false with non-empty skip_reason.
• Forbidden: "write tests later", empty RED for production code, TBD verification.
• Full rules: ../../assets/references/tdd-contract.md.

Task Ownership

If implementer agents are used, each one owns one complete plan task or an explicitly re-planned complete subtask. Do not split one written task into hidden technical slices across workers. Valid reports are DONE, DONE_WITH_CONCERNS, BLOCKED, or NEEDS_CONTEXT.

Ready Gate

Plan is not complete if:

• a PRD acceptance criterion has no mapped task or verification;
• a task has no file target, creation target, or explicit discovery target;
• a tdd.required task is missing RED command, expected failure, or GREEN command;
• verification is vague or non-runnable when a runnable check exists;
• architecture choices are still unresolved and should return to Build;
• task boundaries would cause overlapping write ownership.

Mermaid

Use Mermaid for dependency order or execution flow when text would hide sequencing. Follow ../../assets/references/mermaid-contract.md.