AI Plan

Plan

Decomposes an approved spec into a phased, patch-ready execution plan — the contract /ai-build executes. HARD GATE: operator approves before /ai-build runs.

/ai-plan                    # plan from approved spec
/ai-plan --pipeline=hotfix  # override classification
/ai-plan --skip-design      # skip design routing

Workflow

Principles: §10.6 SDD (approved spec is the contract), §10.3 SOLID (single-concern tasks), §10.5 TDD (RED before GREEN), §10.7 Clean Code (self-review).

1. Approval gate (HARD STOP, no escape hatch) — before decomposing, resolve the spec's CANONICAL state. Read <spec_id> from .ai-engineering/specs/spec.md frontmatter spec: (fallback slug:); run python .ai-engineering/scripts/spec_lifecycle.py status <spec_id>.
   • Sidecar state ≠ approved → HARD STOP, write NO plan.md, emit exactly:

     Error: spec-<id> is in '<state>' state.
     Complete /ai-brainstorm approval before running /ai-plan.
     

   • No sidecar → fall back to spec.md frontmatter status:; block (same HARD STOP) unless status: approved.
   • Neither resolves → indeterminate plumbing (D-161-03): LOUD warning, proceed (fail-open).
   • Vocab: sidecar approved ⇔ frontmatter status: approved. The ONLY bypass is approving the spec via /ai-brainstorm.
2. Read spec — load spec.md; flag missing sections per spec-schema.md.
3. Explore (read-only) — current architecture, patterns, affected files.
4. Classify pipeline — full / standard / hotfix / trivial (table below).
5. Classify executor route — write execution_route frontmatter: executor: build + safe_next_command: "/ai-build" for single-concern, or executor: autopilot + safe_next_command: "/ai-autopilot" for multi-concern/large. status stays the only approval field; drafts are recommendations. Emit framework_operation detail operation=execution_routed.
6. Design routing — invoke handlers/design-routing.md; capture at .ai-engineering/specs/<spec-id>/design-intent.md under ## Design. --skip-design logs reason + proceeds.
7. Identify architecture pattern — read architecture-patterns.md; pick a canonical pattern or ad-hoc; record under ## Architecture BEFORE decomposition.
8. Decompose into tasks — bite-sized (2-5 min), single-agent, single-concern, verifiable, ordered; apply the patch-ready template below (D-131-08 / sub-003).
9. Assign agents — capability-match (build = code; verify = read-only; guard = advisory).
10. Order phases + gates — TDD pairs: a RED test task before any GREEN implementation task.
11. Self-review — spec-reviewer pattern, max 2 iterations.
12. Write + STOP — write .ai-engineering/specs/plan.md, then emit the 6-field value block (Bottom line / Why it matters / What's done / Risk / Next / Details) per .ai-engineering/reference/value-lens.md at the resolved audience level (default full) alongside the safe_next_command, and STOP — operator approves and runs that command. The safe_next_command string itself stays a carve-out (exact, machine-runnable).

Re-plan (plan failed / scope changed, or plan.md holds placeholder content): diff against the existing plan, regenerate affected phases, and preserve completed checkboxes where the task is unchanged.

Output template — patch-ready (D-131-08)

Each task block carries five lines so /ai-build can route mechanical work to the cheap tier:

• - [ ] T-N — <task title>
• - Agent: <build/verify/guard>
• - Files: <path/to/file:line>
• - Principles applied: §10.x ... — cite ≥ 1 CANONICAL §10 anchor (e.g. §10.3 SOLID, §10.5 TDD, §10.7 Clean Code).
• - Patch (deterministic): — a unified-diff hunk when the edit is mechanical (rename, copy, frontmatter add); omit and add prose only when judgment is required.
• - Gate: <test/check>

Routing: patch present → effort: cheap; patch absent or synthesis hint → effort: mid; operator --max-effort → effort: high.

Plan frontmatter MUST include execution_route.version, spec, executor, automation, concern_count, estimated_files, reason, safe_next_command. Do NOT add approved/approval under execution_route; plan status is the approval source of truth.

Pipeline classification

No-execution protocol

/ai-plan is planning-only. MUST NOT invoke ai-build / /ai-build, modify source, or check off implementation tasks. MAY write plan.md and run read-only exploration. Dispatch the ai-plan agent (.claude/agents/ai-plan.md, the interrogator handle) for any approved spec needing decomposition; hand off to /ai-build only after explicit approval.

Common mistakes

• Planning HOW instead of WHAT.
• Omitting the Patch (deterministic): block on mechanical edits — costs /ai-build the cheap-tier dispatch.

Examples

User: "the spec is approved, break it down into a phased plan"

/ai-plan

Reads spec.md, runs read-only exploration, decomposes into phases with agent assignments + gates, writes plan.md, emits the value block, and STOPs for approval.

Integration

Called by: user directly, post-/ai-brainstorm approval. Calls: ai-explore agent (codebase context). Transitions to: /ai-build (only after user approves). See also: /ai-brainstorm, /ai-build, /ai-autopilot (multi-concern alternative).

$ARGUMENTS