Split Into Tasks

SkillProductivity

Split an implementation-ready spec into the fewest independently deliverable vertical tasks and select only risk-justified intermediate review checkpoints. Use when a spec has multiple outcomes needing separate delivery or dependency tracking; keep one coherent outcome as the spec itself.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Split Into Tasks skill

What this skill tells your AI

The instructions your AI receives, as published by toy-crane/claude-code-playbook-template in .agents/skills/split-into-tasks/SKILL.md and read by ahel’s review.

Read the spec

Require docs/specs/<slug>/spec.md; if it is missing, stop before proposing tasks.

Read the spec and current code, plus relevant project decisions, glossary terms, and an approved prototype when present. Treat missing implementation as the target gap. Pause on conflicting settled sources and present the exact decision to resolve. When the approved behavior forms one coherent user workflow, keep the spec as the sole handoff; separate operations or states inside that workflow do not create task boundaries.

Propose the complete outcome map

For work with multiple independently deliverable outcomes, draft the complete shallow map using the fewest independently usable and verifiable outcomes. Fold prerequisite work into the first outcome that makes it useful unless that work is independently deliverable. Declare only blockers that genuinely prevent a dependent outcome.

Cover every spec outcome. For each proposed task show its title, delivered behavior, observable acceptance criteria, focused deterministic verification, task-specific constraints, blockers with reasons, and any review checkpoint. Implementation derives the active outcome's technical approach just in time from the then-current repository, so omit predicted files, functions, code structure, technical-layer steps, context boundaries, and internal sequences. Reference an approved prototype when relevant without copying its implementation.

Add an intermediate review checkpoint only when a material error could compound through substantial dependent work, or when deterministic checks cannot adequately settle a material risk such as security, data, permission, migration, recovery, or an external contract. Name its cumulative scope and concrete risk. A checkpoint buys one review pass over that scope, not a round that repeats until it reports nothing. The implementation phase owns the single final review over the whole diff.

Present the complete proposal and iterate until the user approves it. Write no task files before that approval.

Write the approved handoff

After approval, record shared settled constraints in the spec and only task-specific constraints in task files. Do not invent unsettled behavior.

Write each approved task to docs/specs/<slug>/tasks/<NN>-<slug>.md using templates/task.md, with blockers before dependents. For an approved revision, remove replaced tasks that have never recorded completion. Preserve a task with recorded completion history even if it later returned to in-progress or blocked; after its still-required obligations and blocker references move to the approved replacement, set it to superseded and append revision evidence naming the replacement and reason. A superseded task is terminal for that approved breakdown and is inactive recovery history, not part of the current delivery map. Do not create an archive. End after writing the current task handoff, before implementation.

Signals

GitHub stars
60
Forks
8
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
split-into-tasks
Source
github.com/toy-crane/claude-code-playbook-template