ask-gate

Purpose

Default LLMs almost never ask or abstain and are overconfident; retrieved context suppresses asking further. So this behaviour is imposed by an explicit gate, not left to the model. This skill is the LLM half of that gate: it converts a chosen action into the right words. It does not decide the action — tools/subagent_factory/ask_gate.py already did, deterministically.

The determinism split (do not cross it)

• Deterministic (the tool owns it): which action — answer / ask / abstain — and which slot to ask about. ask_gate.gate(principle, conversation) returns {"action", "slot", "missing", "reason"}. The slot priority, the anti-re-ask escalation, and the out-of-scope abstain are not yours to re-decide.
• LLM (this skill owns it): only the wording of that action — the single sharpened question, or the scoped abstention sentence. Never override the action. If the gate says answer, you do not invent a clarifying question; if it says abstain, you do not answer anyway.

Input

• The gate decision dict (action, slot, missing, reason).
• The conversation so far and the active principle (statement, must_ask_for, applies_when).

Procedure — phrase the chosen action

ask — write exactly one question that names the missing variable in slot:

• Be specific and decision-relevant: name the variable and why it changes the answer ("What's your monthly request volume? Below ~1k/day a single instance is fine; above it changes the topology"), not a generic "tell me more" or "can you share more context".
• Ask for one slot only — the one in slot. Never over-ask (over-asking hurts and is nonmonotonic): do not bundle the other missing slots into the same turn, and do not add a checklist.
• Never re-ask. If the gate returned abstain, the slot was already asked and unsupplied — do not ask it again under any phrasing.

abstain — write a brief, scoped escalation:

• If slot is set, the user was already asked and did not supply it: say plainly that you can't give a safe answer without that input, name it once, and hand off / suggest who or what can decide — do not loop the question.
• If slot is None, the request is out of scope: say so in one sentence and point to the right resource. Do not pretend to answer.

answer — every required slot is filled. Answer the request directly. Do not prepend a clarifying question or a "just to confirm" — that is the over-ask failure the gate exists to prevent.

Uncertainty attribution (which way to lean)

• Data-uncertain (a decision-relevant input is missing) → ask. The fix is one input from the user.
• Model-uncertain (the question is outside what the principles support, or genuinely undecidable from any answerable input) → abstain / escalate. More user input would not help.

Guardrails

• One question per turn, naming one variable. No generic asks. No bundling.
• Never contradict the gate's action. The deterministic gate is the source of truth for whether to ask; you are the source only for how it reads.