Task Planner
Overview
Turn ambiguous or large requests into an executable plan the agent (and user) can follow. Prefer clarity over length.
When to use
- Multi-step projects, migrations, refactors, or "build X end-to-end"
- Unclear scope - plan first, then ask only the blockers
- Handoffs to
spawn subagents or parallel workstreams
When NOT to use
One-shot work is not a plan: one deck, one memo, one file, one script, a typo, a rename, a selected template. Do that now. Do not file a two-step board and wait for Build. HTML or JSON in the pipeline is an intermediate, not a deliverable and not a reason to stop.
Workflow
- Restate the goal in one sentence (success criteria).
- List constraints (deadline, stack, security, “do not touch X”).
- Break into steps (5-12 max for the first pass). Each step must be:
- actionable
- independently verifiable
- ordered by dependency
- Mark dependencies (
blocked by step N).
- Identify risks / unknowns and the cheapest probe to resolve each.
- Define deliverables (files, PRs, reports, commands to run).
- Propose execution mode:
- sequential in this chat, or
spawn for independent tracks (architect / implementer / reviewer)
Output format
Prefer recording the plan on the board + mission ledger. When writing in chat, still include:
## Goal
...
## Constraints
- ...
## Facts
- ...
## Missing info
- ...
## Acceptance criteria
- ...
## Plan
1. ... (validation=verify|test|lint|manual|none; depends on …)
2. ...
## Risks
- ...
## Next action
Start with step 1: ...
Each board step should set acceptance and validation. After filing steps, call board ledger_init only for multi-file / architecture work. A plan that describes a system includes an archify diagram (HTML path + caption).
Rules
- Do not invent requirements - call out assumptions as missing_info or facts.
- Prefer the smallest plan that unblocks progress.
- If the request is a greeting or a vague one-liner with no goal, ask what to plan - do not invent a project name from the word.
- After the user confirms (or in /forge|/cruise with a clear target), execute step 1 unless they ask to wait.
- Wait for Build only after /blueprint on a multi-file or architecture job, or when the user asked only for a plan. A short mechanical plan (1-3 steps) on a simple deliverable starts step 1 in this turn.
- When plans change mid-flight,
ledger_replan locally and state why before continuing.
1---2name: task-planner3description: Decompose complex requests into ordered steps, dependencies, risks, and deliverables. Use when the user asks to plan a project, break down a feature, migrate systems, or tackle multi-step work before coding.4---56# Task Planner78## Overview910Turn ambiguous or large requests into an executable plan the agent (and user) can follow. Prefer clarity over length.1112## When to use1314- Multi-step projects, migrations, refactors, or "build X end-to-end"15- Unclear scope - plan first, then ask only the blockers16- Handoffs to `spawn` subagents or parallel workstreams1718## When NOT to use1920One-shot work is not a plan: one deck, one memo, one file, one script, a typo, a rename, a selected template. Do that now. Do not file a two-step board and wait for Build. HTML or JSON in the pipeline is an intermediate, not a deliverable and not a reason to stop.2122## Workflow23241. **Restate the goal** in one sentence (success criteria).252. **List constraints** (deadline, stack, security, “do not touch X”).263. **Break into steps** (5-12 max for the first pass). Each step must be:27 - actionable28 - independently verifiable29 - ordered by dependency304. **Mark dependencies** (`blocked by step N`).315. **Identify risks / unknowns** and the cheapest probe to resolve each.326. **Define deliverables** (files, PRs, reports, commands to run).337. **Propose execution mode**:34 - sequential in this chat, or35 - `spawn` for independent tracks (architect / implementer / reviewer)3637## Output format3839Prefer recording the plan on the board + mission ledger. When writing in chat, still include:4041```markdown42## Goal43...4445## Constraints46- ...4748## Facts49- ...5051## Missing info52- ...5354## Acceptance criteria55- ...5657## Plan581. ... (validation=verify|test|lint|manual|none; depends on …)592. ...6061## Risks62- ...6364## Next action65Start with step 1: ...66```6768Each board step should set `acceptance` and `validation`. After filing steps, call `board` `ledger_init` only for multi-file / architecture work. A plan that describes a system includes an `archify` diagram (HTML path + caption).6970## Rules7172- Do not invent requirements - call out assumptions as missing_info or facts.73- Prefer the smallest plan that unblocks progress.74- If the request is a greeting or a vague one-liner with no goal, ask what to plan - do not invent a project name from the word.75- After the user confirms (or in /forge|/cruise with a clear target), execute step 1 unless they ask to wait.76- Wait for Build only after /blueprint on a multi-file or architecture job, or when the user asked only for a plan. A short mechanical plan (1-3 steps) on a simple deliverable starts step 1 in this turn.77- When plans change mid-flight, `ledger_replan` locally and state why before continuing.