Do Work
Executes a single ready work item end-to-end. The plan in # Instructions is the contract; this skill follows it rather than redesigning it.
Input
work/<slug>/README.md with status: ready and a populated # Instructions section.
Output
work/<slug>/data/— any data artifacts the plan produced. Each artifact should have a stable name referenced from the README.work/<slug>/README.mdupdated:# Resultspopulated with a natural-language narrative of what was actually done, what artifacts were produced, and anything that diverged from the plan.- Frontmatter
status: pending-assessment.
- A handoff to review-work with the slug.
Procedure
Lock the item. Set frontmatter
status: in-progressto prevent re-entry.Read the plan. Open
# Instructions. If it is empty or vague, setstatus: needs-inputand stop — the plan should have been concrete coming out of plan-work/review-plan. Do not improvise design decisions here.Execute step-by-step. For each step:
- Run the named tool/command/skill with the specified inputs.
- Save artifacts under
work/<slug>/data/using the names from the plan. - If a step fails, capture the error verbatim (do not silently retry with a different approach).
Handle deviations. When reality forces a deviation (a tool produced unexpected output, an external dependency is unavailable, a step turned out to be unnecessary), document it in
# Resultswith: what was planned, what actually happened, why the deviation was justified. Do not silently substitute one step for another.Write
# Results. Structure:# Results ## Summary <one paragraph> ## Artifacts - `data/<file>` — <one-line description> ## Step-by-step <what was done for each plan step, including deviations and failures>Set status and hand off. Set frontmatter
status: pending-assessment, then invoke review-work with the slug.
Out of scope
- Designing or revising the plan (that is plan-work).
- Assessing success (that is review-work).
- Indexing or committing artifacts (that is save-work).