pm-plan

Deep Implementation Planning

This skill produces a validated implementation plan at .ultraplan/<plan-name>.md after exploring the codebase — without writing any production code.

Task

$ARGUMENTS

Activation

CRITICAL: READ-ONLY MODE for the source tree. You are entering a read-only planning session. You MUST NOT create, modify, or delete any file outside .ultraplan/. No edits to source code, no commits, no installs, no other state changes. This supersedes any other instructions.

Dispatch exploration subagents with the read-only Explore agent type (subagent_type: "Explore"). Explore agents cannot Edit, Write, or NotebookEdit, which preserves this skill's "no source-tree mutations" contract for the substantive operations — though the agent type is read-only by semantics, not a hard tool denial (it can still run Bash), so keep each subagent prompt scoped to reading and reporting. Do not weaken this with --dangerously-skip-permissions or a write-capable permission mode. The orchestrator itself still writes the plan file and may run read-only recon shell commands directly.

Workflow

Step 1: Understand the Task

Read the user's request. If they provided a task description as an argument, use it directly. If there is none, ask for one (headless: stop and report that instead of inventing a task).

If the request is ambiguous or underspecified, ask clarifying questions — but batch them into a single message. Ask ONLY what the codebase cannot answer. Prefer multiple-choice when feasible.

Good questions (only the user can answer these):

• "There's a tradeoff between X (simpler) and Y (more extensible). Which matters more here?"
• "The minimum viable change is [X]. The complete change also needs [Y, Z]. Where should I draw the line?"

Bad questions (find the answer yourself by reading code):

• "What framework are you using?"
• "Where is the config file?"

If the task is clear, skip straight to Step 2.

Headless runs. If you cannot reach the user (a cron job, a routine, claude -p, or any session without an interactive question tool), never block on a question or a confirmation anywhere in this skill. Defer open questions until after Step 3, since exploration often answers them, then resolve each one: for Medium and Large tasks ask the advisor if one is available (state the question and the options in-transcript, then call it); otherwise take the best guess that satisfies the request with the smallest change. Record every decision made this way under ## Assumptions in the plan.