Write Executable Plan

Write Executable Plan

Bite-sized plans for parallel agents. No placeholders. No vague steps.

When to use

• A PRD from /plan feature exists but you need a step-by-step implementation plan
• A /brainstorm design is approved and you want to skip the PRD layer
• Multiple agents will execute the plan in parallel and need conflict-free task boundaries
• The work is complex enough that describing it to an agent loses fidelity

When NOT to use

• Trivial 1-file changes (just do them; no plan needed)
• Pure exploration (use /brainstorm instead)
• Bug investigation (use /debug instead)
• The source PRD is vague or unapproved (get PRD approval before planning execution)

Phase 0: Bootstrap Gate

Read skills/_shared/bootstrap-gate.md and execute the gate check. If the gate is CLOSED, invoke skills/bootstrap/SKILL.md and wait for completion before proceeding. If the gate is OPEN, continue to Phase 1.

Phase 1: Source Selection

Accept ONE of the following inputs (in order of preference):

1. PRD path — docs/prd/YYYY-MM-DD-<feature>.md produced by /plan feature
2. Design spec path — docs/specs/YYYY-MM-DD-<slug>-design.md produced by /brainstorm
3. Inline description — user provides a free-text description; prompt for elaboration via AUQ before proceeding

If no source is provided and $ARGUMENTS is empty, ask via AUQ:

AskUserQuestion({
  questions: [{
    question: "Which source should this plan be based on?",
    header: "Plan Source",
    options: [
      { label: "Existing PRD (Recommended)", description: "Point me to docs/prd/YYYY-MM-DD-<feature>.md — most precise decomposition." },
      { label: "Design spec from /brainstorm", description: "Point me to docs/specs/YYYY-MM-DD-<slug>-design.md — skips formal PRD." },
      { label: "Describe inline", description: "Paste or describe the feature — I will ask follow-up questions before planning." }
    ],
    multiSelect: false
  }]
})

Read the source file and extract:

• Feature title (for filename slug and plan header)
• Acceptance criteria or equivalent (drives Task decomposition)
• Explicit out-of-scope items (excludes from plan)
• File inventory if present (seeds the whole-plan Files block)

Phase 2: Decomposition

Break the source into Tasks. Each Task MUST satisfy all of these constraints:

• Single owner — one agent role (e.g., code-implementer, test-writer), not "the team"
• Bounded file set — lists every file it creates, modifies, or tests; no file appears in two Tasks
• Independently testable — after Step 4 the task stands on its own; no cross-task dependencies in the test command
• 2-5 minute wall-clock estimate — if the estimate exceeds 5 minutes, split the Task; if two Tasks share a file, merge or re-scope

Before writing Tasks, output a brief decomposition plan in plain text:

## Decomposition (draft)
- Task 1: <title> — <owner> — ~<N> min — Files: <list>
- Task 2: <title> — <owner> — ~<N> min — Files: <list>
...

If any task's file set overlaps another, surface the conflict and resolve it before writing Steps.

Phase 3: Step Authoring (per Task)

Each Task gets exactly 5 steps in this order. All 5 steps are mandatory.

Step 1: Write the failing test

EARS seam: If the source PRD/spec carries an ## Acceptance Criteria (EARS) (or ## 3.A) section, apply the EARS→vitest 1:1 mapping below to emit per-clause test stubs. If no EARS section is present, fall back to manual derivation from prose.

EARS → vitest mapping (1:1)

Reference test exemplifying this pattern: tests/lib/wave-executor/persona-gate-hook.test.mjs (shipped at #481).

Follow .claude/rules/testing.md § "Test Quality — False-Positive Prevention": one meaningful assertion per it, behaviour not implementation, no branching, hardcoded expected values.

Provide:

• Exact file path (absolute from project root)
• Complete, runnable test code — no // ..., no placeholders, no // implement later
• One sentence explaining why this test verifies the intended behavior

The test MUST fail before Step 3 is applied. If the behavior already exists and cannot fail, pick a harder assertion or a new edge case.

Step 2: Run the test to confirm it fails

Provide:

• Exact command (e.g., npm test -- tests/unit/my-module.test.mjs)
• Expected terminal output — the specific assertion failure line that proves the test exercises missing code (not a generic "FAIL" — copy the actual error shape)

Step 3: Implement the minimal code

Provide:

• Exact file paths labeled Create: or Modify:
• Complete code — every function, every import, every export. No // add appropriate logic, no // similar to above, no ...
• DRY constraint: reuse existing utilities; do not duplicate logic that already exists in the codebase
• YAGNI constraint: implement exactly what makes Step 2's test pass; no speculative features

Step 4: Run the test to verify it passes

Provide:

• Same command as Step 2 (copy verbatim)
• Expected terminal output showing the test suite PASS line with the test name visible

Step 5: Commit

Provide:

• Exact commit message in Conventional Commits format: type(scope): subject
• Scope: the primary module or directory affected by this Task
• Subject: imperative mood, max 72 characters, no period
• Files staged: only this Task's files (list them)

Phase 4: Placeholder Linter

Before writing the plan to disk, scan every Step in every Task for forbidden strings. Any match is a rejection — fix before writing.

Forbidden strings (case-insensitive):

Prose ellipsis (e.g., "Phase 1... Phase 2...") is acceptable. Only code-block ellipsis is forbidden.

Any hit → surface to user via AUQ:

AskUserQuestion({
  questions: [{
    question: "The draft plan still contains forbidden placeholder strings. How do you want to proceed?",
    header: "Platzhalter",
    options: [
      { label: "Fix automatically (Recommended)", description: "I fill in the missing specifics myself, then write the plan — fastest, and you review the result at hand-off.", preview: "[each hit: Task number, Step number, matched text]" },
      { label: "Show me each hit interactively", description: "Walk me through each one so I can provide the missing detail.", preview: "[the same hits, one at a time, starting with the first]" }
    ],
    multiSelect: false
  }]
})

Fix all hits before Phase 5.

Phase 5: Write Plan

Ensure the target directory exists:

mkdir -p docs/plans

Generate today's date via date +%Y-%m-%d. Derive <feature-slug> from the feature title (lowercase, hyphens, no special characters, max 40 characters).

Write to: docs/plans/YYYY-MM-DD-<feature-slug>.md

Use the template in plan-template.md (same directory as this SKILL.md) as the structural guide. Fill every slot. The written plan must:

• Open with # Plan: <feature title> and a Source: / Created: / Status: draft header block. The Source: line MUST carry the tracking Issue/Epic reference inline as #NNN (e.g. Source: docs/prd/2026-07-09-foo.md (#786)). This #NNN is load-bearing: it is the anchor the archive-closed-plans custom-phase reads (via scripts/archive-closed-prds.mjs) to detect that a plan's feature/Epic is closed and archive the plan into the Meta-Vault. A plan with no #NNN in its header is never archived (fail-closed no-epic-ref) — it lingers in docs/plans/ indefinitely.
• Include a whole-plan ## Files section listing all Create / Modify / Test paths across all Tasks
• Contain one ## Task N: <title> section per Task, each with its per-task ### Files block and Steps 1-5
• Pass the Phase 4 linter with zero hits

Phase 6: Hand-off

Present the plan path and task count to the user via AUQ:

AskUserQuestion({
  questions: [{
    question: "Plan written to docs/plans/YYYY-MM-DD-<slug>.md — [N] tasks, ~[total] min. How to execute?",
    header: "Plan fertig",
    options: [
      { label: "Dispatch via wave-executor (Recommended)", description: "Parallel agents execute all tasks simultaneously. Fastest for file-disjoint tasks." },
      { label: "Execute coordinator-direct", description: "One task at a time in this session. Safer for tasks with shared state." },
      { label: "Revise the plan first", description: "I have feedback — describe what to change and I will update the plan." },
      { label: "Done for now", description: "Keep the plan as a reference; no immediate execution." }
    ],
    multiSelect: false
  }]
})

If the user selects "Revise the plan first", incorporate the feedback, update the plan file, re-run the Phase 4 linter, and re-present Phase 6.

If the user selects "Dispatch via wave-executor", hand off to skills/wave-executor/SKILL.md with the plan path as input.

Anti-Patterns

• Steps that say "implement the feature" without exact file paths and complete code — this is the most common failure mode; reject it at Phase 4
• Tasks with overlapping file scopes — agents executing in parallel will conflict; the decomposition must be file-disjoint
• Plans without a per-Task Files block — forces agents to grep the codebase to discover their scope, which loses the speed advantage of planning
• Skipping Phase 4 — placeholders in a plan silently cascade into every agent that executes it; one "TBD" can stall an entire wave
• Estimating tasks at >5 minutes — if a task is that large, it contains multiple logical changes; split it
• Vague commit messages in Step 5 — feat(core): update things is not acceptable; the scope and subject must be derivable from the Task's file set without reading the diff

See Also

• skills/plan/SKILL.md — produces the PRD this skill consumes
• skills/brainstorm/SKILL.md — alternative source (design spec)
• skills/wave-executor/SKILL.md — consumes the plan for parallel execution
• .claude/rules/testing.md § "Test Quality — False-Positive Prevention" — test conventions to follow in Step 1
• skills/write-executable-plan/plan-template.md — structural template used in Phase 5