| name | goal-writer |
| description | Write and create durable, evidence-checkable thread goals when the user explicitly asks to use Goal mode, set a persistent goal, or continue autonomously across multiple turns. Do not use for ordinary one-turn tasks. |
Goal Writer
Turn the user's request into one persistent objective that another turn can resume without relying on conversational intent.
Workflow
- Confirm that the user explicitly requested Goal mode or a persistent multi-turn goal. Never infer Goal mode from task complexity alone.
- Preserve every named deliverable, file, command, Skill, approval gate, release step, and success criterion from the request.
- Write one concise objective containing:
- the concrete outcome;
- the authoritative evidence that proves it;
- constraints and required workflows;
- permission and scope boundaries;
- how to iterate after failures or review findings;
- the condition that genuinely requires stopping for user input.
- Check that the objective is actionable after context compaction and cannot be satisfied by a plausible but unverified final answer.
- Call
create_goal with the objective. Creating a Goal replaces the thread's previous Goal, so do this only for the explicit request currently being handled.
Writing Rules
- State outcomes and evidence, not a speculative implementation plan.
- Keep user wording where it identifies exact artifacts or workflows.
- Include verification appropriate to the task; do not invent unrelated gates.
- Treat approvals, credentials, destructive actions, and external coordination as boundaries rather than permission to bypass them.
- Define blocked as a real impasse requiring user input or an external state change, not as difficulty or a failed first attempt.
- Do not add a Token budget.
- Do not mark the Goal complete. Completion belongs to the executing turns after a strict evidence audit.
Objective Shape
Use this shape as a checklist, not a rigid template:
Achieve <outcome>. Preserve <constraints and boundaries>. Use <required workflow or Skill>.
Verify completion with <authoritative evidence>. If verification or review finds an actionable
issue, fix it and repeat the relevant checks. Stop for the user only when <true blocked condition>.
