goal-decomposition

name	goal-decomposition
description	Trigger — "break this down into tasks", "decompose this goal", "create subtasks for", "plan the work breakdown", "what are the steps to achieve", "split this into agent-sized tasks". Skip — when the goal is already a single concrete action (e.g., "fix typo in README"), when a full task breakdown already exists and is current, or when the user explicitly wants to execute rather than plan.
license	Apache-2.0
compatibility	{"clients":["openai-codex","gemini-cli","opencode","github-copilot"]}
metadata	{"owner":"codex","domain":"goal-decomposition","maturity":"draft","risk":"low","tags":["goal","decomposition","planning","task-breakdown","MECE"]}

Purpose

Transform high-level goals into a tree of agent-executable tasks, each with clear scope, acceptance criteria, and dependency ordering. Ensures every leaf task is small enough for a single agent to complete in one run and that the full tree covers the goal with no gaps or overlaps (MECE).

When to use

• A goal or epic is too large for a single agent to execute in one run.
• The user provides a broad objective ("refactor auth system") that needs structured decomposition.
• A project kickoff needs a work breakdown before agents can be assigned.
• An existing decomposition has gaps, overlaps, or tasks that are still too large.
• Dependencies between tasks are unclear and need explicit ordering.

Do NOT use when

• The goal is already a single concrete action that one agent can complete directly.
• A complete, current task breakdown already exists (update it rather than recreating).
• The user wants execution, not planning — hand off to the appropriate execution skill.

Operating procedure

1. State the top-level goal in one sentence. Write it as a measurable outcome (e.g., "All API endpoints return valid JSON Schema-compliant responses" not "improve the API").
2. Identify the goal's scope boundaries: list what is in-scope and what is explicitly out-of-scope.
3. Decompose the goal into 2–5 sub-goals using MECE principles: each sub-goal is mutually exclusive (no overlap) and collectively exhaustive (full coverage).
4. For each sub-goal, repeat decomposition until every leaf task meets the single-agent rule: completable by one agent in one run (≤30 tool calls, ≤50,000 tokens, ≤300 seconds).
5. Write each leaf task with these fields: | Task ID | Title | Description (≤50 words) | Acceptance Criteria (≤3 items) | Estimated Tool Calls | Dependencies |.
6. Build a dependency graph: for each task, list which other task IDs must complete first. Verify there are no circular dependencies by performing a topological sort.
7. Run the topological sort: list tasks in execution order, grouping tasks that can run in parallel on the same line.
8. Validate MECE coverage: for each sub-goal, confirm that its child tasks fully cover the sub-goal's scope and do not overlap with siblings.
9. Assign priority tiers: Tier 1 = tasks with no dependencies (can start immediately), Tier 2 = tasks depending only on Tier 1, etc.
10. Produce the final Goal Tree as a markdown outline with indentation showing parent-child relationships, plus the execution-order table.
11. Write leaf tasks to the repo's ticket system (e.g., tickets/ directory) if one exists, or output them as a structured list.

Decision rules

• If a leaf task requires more than 30 tool calls, it is too large — split it further.
• If two tasks modify the same file, add an explicit dependency or assign them to the same agent.
• Acceptance criteria must be verifiable by an agent (e.g., "tests pass", "file exists") — never subjective ("code is clean").
• If the goal has ambiguous scope, define the narrowest reasonable interpretation and note assumptions.
• Maximum tree depth is 4 levels. If deeper decomposition is needed, the goal itself is too broad — split the goal first.

Output requirements

1. Goal Statement — one-sentence measurable outcome.
2. Scope Boundaries — in-scope and out-of-scope lists.
3. Goal Tree — markdown outline showing hierarchical decomposition.
4. Leaf Task Table — all leaf tasks with ID, title, description, acceptance criteria, estimated effort, and dependencies.
5. Execution Order — topologically sorted task list with parallel groupings.
6. Dependency Graph — text-based DAG or mermaid diagram showing task relationships.

References

• references/delegate-contracts.md — each leaf task becomes a delegation contract.
• references/checkpoint-rules.md — checkpoints align with sub-goal completion boundaries.
• MECE framework (McKinsey) for ensuring complete, non-overlapping decomposition.

Related skills

• artifact-contracts — each leaf task should have an artifact contract defining its output.
• autonomous-run-control — task sizing must respect agent run budgets.
• human-interrupt-handling — ambiguous goals may require human clarification before decomposition.
• collaboration-checkpoints — parallel task groups need sync points.

Failure handling

• Ambiguous goal: If the goal cannot be stated as a measurable outcome, pause and request clarification — do not decompose a vague goal.
• MECE violation detected: If tasks overlap, merge the overlapping portions into a single task. If gaps exist, create a new task to cover the gap.
• Circular dependency: If topological sort fails, identify the cycle and break it by splitting one task into a dependency-free setup step and a dependent execution step.
• Over-decomposition: If the tree exceeds 30 leaf tasks, re-examine sub-goals for opportunities to merge related tasks.