Nd Task Decomposition

Task Decomposition

Decompose around execution and evidence boundaries, not arbitrary checklist length.

Procedure

1. Define the observable outcome and the end-to-end evidence that proves it.
2. Record the target, required inclusions, negative constraints, dependencies, approvals, and irreversible steps.
3. Split the work into independently verifiable units.
4. Order units by dependency and feedback speed.
5. Choose the smallest reversible first action that produces new information.
6. Expose only Now, Next, and Later unless the user requests the full plan.
7. Re-plan from evidence when a unit fails. Do not repeat the same attempt unchanged.
8. Reconcile every unit and constraint before a terminal claim. A completed unit, build, or lower-layer test proves only itself.
9. Before a long or effectful unit, externalize a restart packet: outcome, last verified state, current action, blocker, remaining authority, and exact re-entry step.
10. On resume, refresh live state and authority before acting. Do not assume volatile context survived or reuse a one-shot approval.

Check the task for:

• initiation: the first action is unclear or costly
• working memory: important state exists only in conversation
• sustained attention: a unit runs too long without feedback
• planning and priority: dependency or value order is unclear
• time management: no duration range, checkpoint, or stop condition exists
• flexibility: the plan has no failure route
• metacognition: done cannot be observed

Modify the task or environment before demanding more effort.

Output Shape

Outcome: ...
Now: one concrete action
Next: one dependent action
Later: deferred work
Re-entry: exact restart step and remaining authority
Done when: executable evidence

Source Basis

Synthesize Dawson and Guare's executive-skills taxonomy and intervention sequence: assess fit, modify the environment or task, teach or support the needed skill, externalize steps, review evidence, and fade scaffolding when it is no longer useful.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jpoindexter
• Source: jpoindexter/nd-skills
• License: MIT

Install and usage instructions live in the source repository linked above.