Grill Me Quick

Apply the grill-me skill to the user's target.

for each grill-me question:
  decision point?
    ├ no  → continue
    └ yes → draft answer (tag each claim as `load-bearing` or `peripheral`)
              ↓
            self-score two axes (1–10):
              • fact   — are the claims true?
              • reason — is the argument sound for this context?
              ↓
            confidence = min(fact, reason)   # if fact is N/A, use reason alone
              ↓
            escalate to subagent verification if:
              • confidence is borderline (5–7) AND draft has factual claims, OR
              • high blast radius AND draft has factual claims
              ↓
            (if escalated)
              dispatch one subagent per claim, in parallel
              → aggregate verdicts → update `fact`
              → recompute `confidence = min(fact, reason)`
              ↓
            passes gate?
              ├ yes → lock in; surface {question, alternatives, scores+rationale, answer}
              │         if escalation fired:
              │           any deviation (refuted/uncertain/re-tagged)
              │             → include `verification:` block listing only deviations
              │           otherwise
              │             → add "N claims verified" to rationale
              └ no  → ask user

Claim tagging

Each factual statement in the draft must be tagged as either:

• load-bearing — removing it would invalidate the decision.
• peripheral — supporting detail; the decision still holds without it.

Tagging is identification. An untagged sentence is not a claim and will not be verified. Do not tag subjective or hedged statements ("team is familiar", "usually works well").

Claim granularity

• Load-bearing claims must be atomic: each one verifiable on its own. If a subagent would have to branch ("true for X, false for Y"), split first.
• Peripheral claims may be composite when natural to state, as long as each composite is verifiable end-to-end by a single subagent.

Heuristic: if stating it takes more than one "and", it is probably two claims — atomize if load-bearing.

Escalation

Default: self-score both axes and let the gate decide. This keeps the skill quick. Escalate only when self-score is borderline (5–7), or blast radius is high with factual claims present (even at 8–10).

Accepted tradeoff: if the main agent is overconfident AND wrong on a factual claim (Dunning–Kruger), the escalation path will not catch it. For quick mode this is acceptable.

When escalated, follow references/verification.md for the subagent pack, output schema, and aggregation rules.

Confidence gates

• 8–10 → lock in.
• 5–7 → lock in if reversible and low blast radius; otherwise ask user.
• 1–4 → ask user.

When to escalate

Situation	Action
confidence 5–7 AND draft has factual claims	Escalate
High blast radius AND draft has factual claims	Escalate (even at 8–10)
Pure judgment, no factual claims	Skip; fact = N/A, decide on reason alone
confidence 8–10 AND low blast radius	Skip, lock in
confidence 1–4	Skip, ask user