| name | starting-task |
| description | Use when beginning a new piece of work and deciding whether to create a tracked task in `.my-context/tasks/`, especially before research, planning, or implementation on non-trivial work. |
Starting Task
Use this skill when new work arrives and you need to decide whether it should become a tracked task.
Purpose
This skill decides when durable task tracking is worthwhile. pi-task provides direct task-file operations; context-task defines document-quality policy.
When To Use This Skill
Use this skill when:
- the user asks to start or track a task
- a request is substantial enough to benefit from durable tracking
- the work will likely need research, planning, review, or multiple sessions
Usually do not use it for tiny one-shot edits, purely informational questions, or throwaway exploration.
Decision Rule
Create a tracked task when one or more of these is true:
- the work is multi-step or spans multiple files
- the work may take multiple sessions
- the work benefits from explicit notes, planning, or completion history
- the user explicitly asks to track it
If none apply, proceed without task initialization.
Start The Record
Use task_start to create the record. Supply a concise title and all stable information known from the request: ID when one is supplied, ticket reference, focused goal, concrete acceptance criteria, and created date when known.
Prefer a safe inferred first draft over unnecessary follow-up questions. Generate a concise title, focused goal, and concrete acceptance criteria when the user did not provide them; ask only when the request is too ambiguous for a safe default.
task_start creates the active task record. Do not create task folders, edit state.json, or manipulate task paths directly.
After Initialization
For tracked work, use the natural phases:
- Research — use
context-task and context-memory; record task-specific findings with task_append_note.
- Plan — use
writing-plans; persist the coherent plan with task_write (document: "plan").
- Execute — use
executing-plans or subagent-driven-development.
For simple, well-scoped work these phases may happen in one session. For larger or riskier work, pause at natural checkpoints and confirm with the user before proceeding.
Never do these automatically:
- Create or raise a PR without explicit user instruction.
- Close an external ticket (GitHub issue, Jira, Linear, and so on) without explicit user instruction.
- Finish a task with
task_finish without explicit user instruction.
Red Flags
Do not:
- create tracked tasks for every trivial request
- ask for fields that can be inferred safely
- manipulate task folders, task documents, or
state.json directly
- create or raise a PR without explicit user instruction
- close an external ticket without explicit user instruction
- finish a task without explicit user instruction
Bottom Line
Use starting-task to decide whether work deserves durable tracking, then use task_start to create its active record with the best available stable brief.