validate-answer

name

validate-answer

description

Trust-check an answer, plan, or decision you already have — multiple identical independent reads converge per question, agreement signals confidence and splits flag what's contested, then walk each decision. TRIGGER when: user says 'validate', 'trust-check', 'is this right', 'pressure-test my decision'; user has a near-final answer or a few focused decisions to confirm. SKIP when: no answer yet, want one built (best-answer); surfacing what's missing (find-gaps).

Validate Answer

Trust-check an answer or a few focused decisions.

When to use

You have an answer, plan, or decision and want to know if you can trust it.

• Multiple identical independent reads: agreement = confidence, disagreement = contested. The reviewers are deliberately the same — diversity here would destroy the signal.
• NOT this if you have no answer yet and want one built → best-answer.
• NOT this if you want to surface what you missed → find-gaps.
• NOT this if you're grading one concrete proposal → second-opinion.

Steps

Step 1 — Dispatch

• Reviewers: R0 (you, concurrent) + R1, R2 (two general-purpose subagents in parallel).
• Prompt (identical for all three): artifact path(s), the questions, a confidence (0.00–1.00) and impact score per recommendation.
• Preflight: if any question is vague, sharpen via AskUserQuestion first.
• Output schema per question: recommendation, confidence (0.00–1.00), impact — render Label (value): Minimal (0.25) · Low (0.5) · Medium (1) · High (2) · Massive (3) — 1-2 sentence reason.
• Unprompted observations (any reviewer): 1-sentence claim, raising reviewer(s), confidence 0.00–1.00, impact; merge duplicates.

Step 2 — Classify, synthesize, and confirm

• Classify each question by its three reviewer scores — here unanimous agreement is the trust signal:
  • agreed — all three ≥ 0.75. Unanimous across identical reviewers; the answer holds.
  • contested — ≥1 reviewer ≥ 0.75 OR ≥2 reviewers ≥ 0.70. Real support, not consensus — walk it, flagged as contested.
  • drop — ≤1 reviewer ≥ 0.70 and none ≥ 0.75. Too thin to walk.
• Order: sort the walk set by max score descending.
• Dependency override: if a decision can invalidate later questions, walk it first.
• Table:

| # | Question | R0 | R1 | R2 | Impact | Band |

Band = agreed or contested.

• Checkpoint: use AskUserQuestion to confirm the walk order before walking.

Step 3 — Walk questions one at a time

Queue rule: drop later questions invalidated by an approved decision, with a one-line reason.

For each question:

• Present: restate the question, quote the artifact span if location-specific, show R0/R1/R2 side-by-side, and say whether it came in agreed or contested.
• Decide: AskUserQuestion with options: apply / alternative / defer. Apply if actionable; otherwise record.
  • On pushback: invoke the second-opinion skill via the Skill tool on the challenged pick.

Step 4 — Walk unprompted observations

List all observations; walk those at confidence ≥ 0.80 (Step 3's sub-procedure, skip R0/R1/R2 split) — a solo observation is uncorroborated, so it clears a higher bar than a voted question.

Step 5 — Summary

One line each:

• Decisions applied — question # + pick.
• Decisions deferred — reason.
• Questions skipped (upstream invalidation) — reason.
• Questions dropped (Step 2 drop band) — question + score.
• Unprompted observations — applied / deferred / dropped.