Project Decomposer
Turns a project into tasks somebody can actually finish, each with a stated way to tell it is done.
This skill does no work. It decomposes and stops. If a task it produces looks like something
Cowork could do, that is a separate request the user makes deliberately.
Inputs
- A project description from the user, or a named document found through Enterprise Search
- Existing context: prior decisions, related threads, the people already involved
references/config.md for team names, the default task size and the house verification bar
Procedure
- Read the goal. Restate it in one sentence. If the restatement is longer than one sentence,
the project is two projects: say so and ask which one to decompose.
- Find the real constraints. Search mail and files for stated dates, named owners, approvals
already given, and anything previously tried. Decomposing against invented constraints is worse
than not decomposing.
- Break into tasks at the level where one named person could finish one task in the default
size from config. Not phases. Not workstreams. Tasks.
- For every task, write all five fields. A task missing any of them is not a task yet:
- What happens, in a verb phrase
- Owner, a named person or a stated vacancy. Never "the team"
- Done when, the verification criteria: what somebody checks to agree it is finished
- Blocked by, naming the task or decision it waits on, or "nothing"
- Evidence, the link to where the requirement came from, or "assumed" if you inferred it
- Order by dependency, then write the critical path in one line.
- Return one Adaptive Card per task, plus a summary card carrying the goal, the task count,
the critical path and every open vacancy.
The verification bar
"Done when" is where most decomposition fails. These are not acceptable:
- "When it's finished"
- "When the team is happy"
- "When reviewed"
These are:
- "When the signed PDF is in the contracts library"
- "When the test suite passes on main and the owner has posted the run link"
- "When three named stakeholders have replied approving in writing"
If you cannot write a checkable "Done when" for a task, the task is still too vague. Split it or
say plainly that it needs a decision before it can be planned.
Output
One summary Adaptive Card, then one card per task. Also write the whole set to
/Documents/Cowork/output/decomposition-<project>-YYYY-MM-DD.md so it survives the session.
Definition of done
Every task has all five fields, every "Done when" is checkable by someone other than its owner,
and the critical path is one line.
Returning fewer tasks is usually better. A decomposition with 9 real tasks beats one with 30
where half are "coordinate" and "follow up". If the project genuinely resists decomposition, say
what is missing rather than padding the list.
When there is no data
- The goal is unclear: ask one question and stop. Do not decompose a guess.
- No owners can be identified: write the vacancies explicitly as "OWNER NEEDED" and count them in
the summary. A plan with named gaps is useful; a plan with invented owners is not.
- No dates found anywhere: say so. Do not invent a schedule.
Boundaries
- Plans only. Never create tasks in Planner or To Do, never send anything, never book time.
- Never assign a named person who has not already been associated with this work in the evidence.
Proposing an owner is fine when it is labelled as a proposal.
- Never state a date as agreed unless the evidence shows it was agreed.
Self-check before returning
- Does every task have all five fields?
- Could a stranger check every "Done when" without asking the owner?
- Is every owner evidenced, or labelled as proposed, or marked OWNER NEEDED?
- Did I avoid creating anything, and only plan?
1---2name: project-decomposer3description: Breaks a project, initiative or goal into owned tasks, each carrying its own verification criteria and a named owner, and returns one Adaptive Card per task. It plans and does not execute. Use when the user asks to break down a project, decompose an initiative, turn a goal into tasks, work out who does what, or says a project is too big to start.4metadata:5 source: https://skillmd.com/skills/kesslernity/project-decomposer6---78# Project Decomposer910Turns a project into tasks somebody can actually finish, each with a stated way to tell it is done.1112**This skill does no work.** It decomposes and stops. If a task it produces looks like something13Cowork could do, that is a separate request the user makes deliberately.1415## Inputs1617- A project description from the user, or a named document found through **Enterprise Search**18- Existing context: prior decisions, related threads, the people already involved19- `references/config.md` for team names, the default task size and the house verification bar2021## Procedure22231. **Read the goal.** Restate it in one sentence. If the restatement is longer than one sentence,24 the project is two projects: say so and ask which one to decompose.252. **Find the real constraints.** Search mail and files for stated dates, named owners, approvals26 already given, and anything previously tried. Decomposing against invented constraints is worse27 than not decomposing.283. **Break into tasks** at the level where one named person could finish one task in the default29 size from config. Not phases. Not workstreams. Tasks.304. **For every task, write all five fields.** A task missing any of them is not a task yet:31 - **What** happens, in a verb phrase32 - **Owner**, a named person or a stated vacancy. Never "the team"33 - **Done when**, the verification criteria: what somebody checks to agree it is finished34 - **Blocked by**, naming the task or decision it waits on, or "nothing"35 - **Evidence**, the link to where the requirement came from, or "assumed" if you inferred it365. **Order** by dependency, then write the critical path in one line.376. **Return one Adaptive Card per task**, plus a summary card carrying the goal, the task count,38 the critical path and every open vacancy.3940## The verification bar4142"Done when" is where most decomposition fails. These are not acceptable:4344- "When it's finished"45- "When the team is happy"46- "When reviewed"4748These are:4950- "When the signed PDF is in the contracts library"51- "When the test suite passes on main and the owner has posted the run link"52- "When three named stakeholders have replied approving in writing"5354If you cannot write a checkable "Done when" for a task, the task is still too vague. Split it or55say plainly that it needs a decision before it can be planned.5657## Output5859One summary Adaptive Card, then one card per task. Also write the whole set to60`/Documents/Cowork/output/decomposition-<project>-YYYY-MM-DD.md` so it survives the session.6162## Definition of done6364Every task has all five fields, every "Done when" is checkable by someone other than its owner,65and the critical path is one line.6667**Returning fewer tasks is usually better.** A decomposition with 9 real tasks beats one with 3068where half are "coordinate" and "follow up". If the project genuinely resists decomposition, say69what is missing rather than padding the list.7071## When there is no data7273- The goal is unclear: ask one question and stop. Do not decompose a guess.74- No owners can be identified: write the vacancies explicitly as "OWNER NEEDED" and count them in75 the summary. A plan with named gaps is useful; a plan with invented owners is not.76- No dates found anywhere: say so. Do not invent a schedule.7778## Boundaries7980- **Plans only.** Never create tasks in Planner or To Do, never send anything, never book time.81- Never assign a named person who has not already been associated with this work in the evidence.82 Proposing an owner is fine when it is labelled as a proposal.83- Never state a date as agreed unless the evidence shows it was agreed.8485## Self-check before returning8687- Does every task have all five fields?88- Could a stranger check every "Done when" without asking the owner?89- Is every owner evidenced, or labelled as proposed, or marked OWNER NEEDED?90- Did I avoid creating anything, and only plan?