Turboplan: Shell Mode

Fill-in → refine → confirm → implement for a shell file path that was passed.

Task Tracking

Use TaskCreate to create a task for each phase:

Run the /draft-plan skill with the shell file path as input. The resulting plan path equals the shell path.

Run the /refine-plan skill, passing the filled-in plan path. Loops until the plan stabilizes.

Report the plan's location and summarize it in one paragraph. Point the user at the file path so they can open it for full detail.

Use AskUserQuestion to ask how to proceed:

Implement now — Record the choice and proceed to Phase 4.
Revise more — Use a follow-up AskUserQuestion to choose: run the /refine-plan skill again (if review-loop refinement is needed) or the /draft-plan skill again (if the shape itself needs rework). After the chosen skill completes, return to Phase 3 and re-run this confirmation gate.
Stop — Record the choice and proceed to Phase 4. Leave the plan file in place.

Run the /self-improve skill to compound planning learnings.

Implement now — Run the /implement-plan skill, passing the plan path. The plan's final implementation step should mark the corresponding prompt done in the prompt plan index.
Stop — Halt with this message:

Plan ready at <plan path>.

Planning context is likely full, and the plan is comprehensive enough to continue fresh. When you're ready to implement, run /clear, then /implement-plan with the plan path.

Phases 1, 2, and 4 always run. Phase 4 runs before the Phase 5 branch so learnings are captured even if implementation never completes.
The AskUserQuestion gate in Phase 3 is mandatory. Never chain into implementation without explicit user approval.
Do not edit the plan file directly. Revisions go through /refine-plan or /draft-plan.
If the user keeps choosing "Revise more" more than twice, surface the pattern and ask whether the shell itself needs to be re-scoped.