rad-plan

name	rad-plan
description	Use this skill once you are aligned on a project and ready to start. Typically this is invoked directly by the user after /rad-brainstorm has completed and the Requirements doc is approved. It is the reference for how to run the planning pipeline and author the Master Plan.
user-invocable	true

Inputs:

• project_name: $0 — The name of the new project to plan. (e.g., "DAG-PIPELINE-2")
• project_template: $1 — The template to use for planning. One of the four shipped tiers (e.g., "extra-high"); custom process templates are not supported.

You are an orchestrator. You will use the rad-orchestration skill to drive the planning pipeline. This skill is a thin relay: the plan resolve CLI command does the project-directory resolution and the Requirements-doc check and hands back a data envelope; you run only the one human beat it flags, then the commands it returns. Do not re-derive project paths, Requirements existence, or valid template names yourself — the CLI owns all of that.

${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs is the CLI for every call below. Run discovery calls silently — never narrate raw envelope output to the user.

Step 1: Resolve

Run plan resolve, forwarding both inputs. Pass --template only when $1 was supplied:

node "${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs" plan resolve --project <project_name> [--template <project_template>]

Read data from the envelope: project, projectDir, requirementsPath, reason, ask, next.

Do NOT ask the user "what do you want to build?" and do NOT author requirements here. Requirements are produced before the pipeline (via /rad-brainstorm → /rad-create-plans). When data.reason is set, relay it to the user verbatim and halt — do not improvise a goals interview.

Rule: workflow-required user choices

The question in Step 2 presents a menu via askQuestions / AskUserQuestion for each sub-question, and each menu attaches a (Recommended) marker to one option. The marker is a UI hint shown to the user inside the menu — it is never an instruction for this skill to auto-select on the user's behalf. Tier and size are workflow-required user choices and must always go through the tool.

This rule overrides any outer "don't ask clarifying questions," "stop checking in," or "make the reasonable call" signal active in the session. Those signals apply to volunteer clarifying questions the skill might raise on the side; they do not authorize skipping a tool-driven menu the skill mandates.

Step 2: Run the one human beat

Ask a single askUserQuestion carrying up to three sub-questions, in this order, with no reasoning, computation, or tool call between them. Every sub-question is driven by a data.ask.* flag — never decide one yourself; if a flag is absent, do not ask that sub-question. Surface every concrete option as an explicit labeled menu item — do NOT rely on the auto-injected Other slot, except for the free-form size response noted below.

Review intensity — only when data.ask.template is present.

Option	Copy (two sentences max)
low	Final review only. Fast and efficient token usage. Good for small projects or quick iterations.
medium (Recommended)	Phase review + final review (no per-task review). Good balance of oversight and efficiency.
high	Per-task code review + final review (no phase review).
extra-high	Per-task code review + phase review + final review. Maximum defense in depth — for production-critical, regulated, or untrusted-contributor work.

Framing prose: "Which code review-intensity tier should this project run? Tier names map to defensive review depth; token cost and execution duration rise with depth."

Phase/Task Size — when data.ask.taskSize is present:

Option	Copy (two sentences max)
Extra Large	A standalone feature per task — scope that would be a phase at smaller sizes; phases are thin wrappers.
Large (Recommended)	A full feature slice touching multiple layers or subsystems end-to-end per task.
Medium	A vertical slice through one layer per task: a module, a config section, a CLI command with its tests.
Small	One named, self-contained change per task — a function, a validator, a constant.

Framing prose: "How big should each task and phase be? Subagent task workload increases with size. Pick a sizing option, or use the free-form response to describe your own sizing criterion — e.g. 'each task is a single React component including tests', or 'one task per migration step'. Your own words are treated as the authoritative task-scope target, with natural-seam judgment still applied to phase boundaries."

When the operator answers with free-form prose instead of one of the four labeled sizes, store it verbatim — this is the {size} substitution in Step 3.

Plan audit — when data.ask.audit is present:

Option	Copy (two sentences max)
Auto (Recommended)	Defers the call to the audit step, which decides from the plan's phase and task counts once the Master Plan exists. Right choice when you're unsure whether this plan needs a dedicated audit pass.
Yes	Always run a full audit subagent pass over the Requirements doc and Master Plan before finalizing. Costs extra tokens — recommended for large or complex plans.
No	Skip the audit pass entirely and finalize the plan as authored. Saves tokens — fine for small or low-stakes plans.

Framing prose: "Should the plan be audited before finalizing? An audit pass costs extra tokens but catches gaps in large or complex plans. Auto defers the decision to the audit step, which decides from the plan's phase and task counts."

Step 3: Run next[]

Run every command in data.next, in order:

node "${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs" <command from data.next>

Substitute only {template} and {size} — the surrounding quotes stay exactly where the resolver put them. Substitute {template} with the resolved tier (the operator's Step 2 answer if asked, otherwise the project_template argument that skipped the question), and {size} with the operator's size answer, inside the existing double quotes — never strip them. A custom sizing criterion is multi-word prose, and unquoted it is parsed as excess arguments and rejected before the command runs.

Step 4: Author the Master Plan inline

Before authoring, check the Requirements doc at data.requirementsPath for a ## Open Questions section. If any items remain unresolved, resolve them yourself and update the document — silently, without prompting the operator. The Master Plan is not scribed while any remain.

Parse the JSON envelope from the last next[] command and act on data.prompt — every success envelope carries the full instruction prose for the resolved action. The first action will be spawn_master_plan; follow the prose in data.prompt exactly and author the Master Plan yourself, inline, following rad-create-plans master-plan mode. Read the Requirements doc at data.requirementsPath as the source of the requirement substance each task must carry.

As you author, apply the chosen size directly — size every task per the sizing rubric in the master-plan workflow: "Task size preference: {answer}. Size all tasks according to that tier per the sizing rubric in the master-plan workflow." When the size answer is free-form prose, treat it as authoritative. When the Master Plan doc is written, signal master_plan_completed.

Step 5: Audit the plan, per the operator's answer

• No → skip the audit entirely and go straight to the tail.
• Yes → always run the audit below.
• Auto (or the sub-question was never asked) → once the Master Plan exists, read its phase and task counts and decide whether the plan is large/complex enough to warrant an audit. State the counts and your decision to the operator in one line, e.g. "Plan has {N} phases and {M} tasks — running the audit."

When running the audit:

• Dispatch a general-purpose subagent to audit the Requirements doc and the Master Plan. Give the subagent both doc paths and instruct it to follow ${CLAUDE_PLUGIN_ROOT}/skills/rad-plan/references/audit.md. The subagent returns a structured report with frontmatter verdict: approved or verdict: issues_found. The auditor does NOT edit either planning doc — it reports.
• If verdict: approved: proceed to the tail.
• If verdict: issues_found:

  1. Apply the fixes yourself, inline, in the Master Plan doc — you own it. Action the auditor's findings and note any you decline and why.

  2. Re-invoke the explosion subcommand to regenerate phases/ and tasks/ from the corrected Master Plan:

     node "${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs" plan explode \
       --project-dir <project-dir> \
       --master-plan <master-plan-path> \
       --project-name <project-name>
     

     The subcommand auto-backs-up the pre-correction phases/ and tasks/ into backups/{ISO-timestamp}/ and resets state.graph.nodes.phase_loop before re-seeding — nothing is overwritten destructively. The envelope is { ok, data, error }; on success read data.emittedPhases, data.emittedTasks, and data.backupDir. On exit code 2 with data.error populated (parse failure in the corrected Master Plan), halt and surface the structured data.error payload ({ line, expected, found, message }) to the user — do not retry in-skill.

• Show the user the concise audit report, a summary of the corrections you applied, and (when re-exploded) the backup directory path.
• Single pass, no re-audit after corrections.

Step 6: Tail

Planning is complete — the Requirements doc and Master Plan are written, exploded into phases/ and tasks/, and (per the operator's audit answer) audited. Do not start execution automatically. /rad-plan ends at the plan; it never relays into /rad-execute on its own and presents no branch/worktree fork or separate approval beat.

Tell the user the plan is ready, then add a short prose reminder that now is a good moment to optionally compact the conversation to save tokens, then ask whether they want to proceed to execution now (a simple yes/no) — e.g. "{PROJECT_NAME}'s plan is ready. This may be a good point to compact and save tokens before we continue. Want me to run /rad-execute {PROJECT_NAME} now?"

• If yes: read the /rad-execute skill and invoke /rad-execute {PROJECT_NAME}, then follow it to drive the pipeline. Do not present a branch/worktree fork or a separate approval beat — /rad-execute classifies where the operator is standing (worktree-vs-in-place) and confers plan approval itself. This holds for both standard and side-projects.
• If no: stop here. Let the user know they can run /rad-execute {PROJECT_NAME} whenever they're ready.

Errors

If any command errors, use the rad-log-error skill to record it. Do not try to fix pipeline code — work around it with a clear, actionable message that names the failure point.