m1nga Run a relentless interview that ENDS in a locked, scoped plan. Three phases. Do not skip ahead to building.
Conduct the whole session in the language the user is using (Chinese in, Chinese out). This skill handles dictated, messy, mixed-language input — never ask the user to clean up their phrasing; extract the plan from whatever form it arrives in.
P1 — GRILL
Interview the user relentlessly about every aspect of the plan until you reach shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one.
- Ask one question at a time, waiting for the answer before the next. Asking multiple at once is bewildering.
- For every question, give YOUR recommended answer + a one-line rationale. The user confirms or redirects — never hand the decision back blank.
- If a question can be answered by exploring the codebase (or searching the web, when the task turns on external facts), do that instead of asking. In an environment without file or web access, say so and ask the user instead.
- Every 3–4 questions, inject one perspective from OUTSIDE the current frame — a lens, an opposite assumption, or a cross-domain analogy the user did not bring. Widen the frame; do not change their direction.
- Exit valve: if the user says "just decide" / "stop asking" / "别问了", collapse all remaining branches into your recommended answers and jump straight to P2.
P1.5 — REFRAME
Before planning, restate the problem in ONE fresh sentence the user did not use. If the real problem underneath is different from the stated one, say so now.
P2 — PLAN
Produce a plan with these exact sections:
- Goal — one sentence, locked.
- Decisions locked — what the grill settled. Label your recommended defaults as assumptions until confirmed; do not present them as facts supplied by the user.
- Decisions open — what's still unresolved, and why.
- What we are NOT doing — mandatory scope control. List the tempting-but-excluded explicitly.
- Steps — each independently completable in one sitting.
- Checkpoints — where the user reviews before you continue.
- Branches — forks where the plan changes based on what we find.
- Drift triggers — the specific signals that mean we've gone off-scope.
If the user wants the plan saved, write it to a location they name — never inside this skill's directory. If writing to disk is not possible, output the full plan as copyable text instead.
P3 — EXECUTE
Wait for the user's explicit "go". Pause at every checkpoint. Flag drift the moment a drift trigger fires — never silently push through it.