task
Use this skill when the user wants to plan a new feature, bug fix, enhancement, documentation update, or chore before implementation. Creates or updates a reviewed task document in docs/task/ and updates TASKS.md, then stops for human approval.
How do I install this agent skill?
npx skills add https://github.com/eljun/claude-workflow-skills --skill taskIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill facilitates a structured planning workflow by reading project documentation and researching the codebase to generate task plans. It is generally safe due to a mandatory human-in-the-loop review step and the absence of code execution or network operations, though it remains theoretically susceptible to instructions embedded in the project data it processes.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
task - Planning Stage
Invariants
- Create or update the task document only.
- Do not implement code.
- Do not start automation from planning.
- Stop after updating
TASKS.mdand tell the user the task is ready for review. - After adding the row to
Planned, re-readTASKS.mdto confirm the row is present before handoff. - Never change the scope of a task that is already claimed or
In Progress(a worker branch, an open PR, or the tracker says so). Put new findings in a new task that references it, and tell the user to message the agent working on the original. Editing a claimed doc does not reach an agent that already read it.
Runtime adapters may expose this stage as a slash command, menu action, or natural-language skill invocation. The portable stage name is task.
Workflow
- Read
AGENTS.md. - Read
LEARNINGS.mdif it exists. - Clarify the request only where missing information changes scope, risk, or acceptance criteria.
- Perform targeted codebase research for files directly relevant to the task.
- Coordination check. List the files this task will change, then read the other
PlannedandIn Progresstask docs (and open worker branches or PRs, if the project uses them). Record every overlap inCoordination, naming the shared file and the other task, with an ordering rule ("land after {ID}") when two tasks edit the same file. - Open decisions. Anything only the user can decide goes in
Open Decisions, markedBLOCKING(implementation must not start) orNON-BLOCKING(a default is stated). Don't bury a decision in prose. - Operator steps. If merging the change does not by itself apply it (migrations, deploys of separately deployed components, config, data backfills), list those steps in
Operator Steps. For deploys, name every affected deployable, including those that only import a changed shared module, and say how you found them. - Choose a recommended execution tier.
- Create or update
docs/task/{ID}-{task-name}.mdfrom the template (see Templates). - Add or update the task in
TASKS.mdunderPlanned. - Stop for user review and approval.
Templates
Use the project's own docs/templates/task-document.md and docs/templates/tasks-tracking.md when they exist. Otherwise use the copies bundled with this skill in templates/ (next to this file). If the project's existing task docs follow a different shape, match the project, but keep the required sections below.
Execution Tier Guidance
Use Recommended Tier in the task document:
| Tier | Use When |
|---|---|
fast | Small, low-risk changes that follow existing patterns. |
balanced | Normal feature work, moderate refactors, or changes across several files. |
deep | Architecture, security, data model, migration, high-risk refactor, or ambiguous requirements. |
The tier is guidance for downstream agents. Runtime adapters define how, or whether, tiers map to models or reasoning settings.
Task Document Requirements
The task document must include:
- clear overview and requirements
- explicit out-of-scope or must-not-change boundaries
- proposed file changes
- focused code context for files the implementation stage should read
- implementation steps
- acceptance criteria
- verification commands or approach
Coordination(overlapping tasks and files, or "None found")Open Decisions(eachBLOCKINGorNON-BLOCKING, or "None")Operator Steps(or "None: merge is enough")- compatibility touchpoints for packaging, docs, adapters, or install surface when relevant
Acceptance criteria are checks, not prose. Each one names the command, query, API call or browser step that verifies it, the expected result, and whether it runs pre-merge or post-deploy. Include at least one no-regression check for behaviour that must not change. A criterion with no check that could fail can't be claimed later; if it can't be automated, mark it manual and say why.
When the task comes from a live investigation, include the evidence (the query, record IDs, counts, file:line) so the implementation agent can re-verify it instead of trusting prose.
Embed concise code excerpts or notes in Code Context; do not paste entire unrelated files.
TASKS.md Update
If TASKS.md does not exist, create it from the tracker template (see Templates).
If it exists, match its existing sections and columns, even if they differ from the template. Add the task to Planned (or the project's equivalent), for example:
| {ID} | {Task} | {Priority} | {Type} | docs/task/{ID}-{task-name}.md | {Date} |
If the task has a BLOCKING open decision, say so in the row ("⚠️ decision D1 needed before implement").
Handoff
End with:
Task planned: docs/task/{ID}-{task-name}.md
Status: awaiting user review
Blocking decisions: {list or "None"}
Coordinates with: {task IDs or "None"}
Next stage after approval: implement {ID}
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/eljun/claude-workflow-skills/task">View task on skillZs</a>