Find Gaps

Surface what an in-progress artifact is missing — many parallel lenses, scored findings, walked one at a time.

When to use

You have a draft, answer, or code and want to surface what you missed — blind spots, risks, overlooked angles.

• NOT this if you want one synthesized answer → best-answer.
• NOT this if you want a yes/no trust signal on an answer you have → validate-answer.
• NOT this if you're grading one concrete proposal → second-opinion.

Steps

Step 1 — Frame and derive lenses

• Derive lenses from the artifact's risk surface — orthogonal angles that together cover what could be wrong or improvable (for a skill: resumability, failure-modes, coherence-with-siblings, UX; for a spec: scope, architecture, edge cases). One lens per subagent; merge near-duplicates — overlap re-raises one finding and wastes the fan-out. Every run also carries one standing lens: unstated assumptions — what the artifact treats as settled but never justifies.
• Count: default 3–5 lenses; scale up only for a broad audit. If the user named the exact lenses, honor them and skip the checkpoint; if partial, honor those and fill any gap.
• Checkpoint: print the lens list (one line each) and confirm via AskUserQuestion ("Run these" / "Adjust").

Step 2 — Dispatch

• Readers: R0 (you, concurrent) hunts what no derived lens covers — cross-cutting gaps and whole categories the lens list missed — and returns findings under the same return contract; R1…RN are parallel general-purpose subagents, one focused lens each.
• Prompt (per lens): artifact path(s), the lens and what it must scrutinize, and the return contract.
• Return contract: finding · recommended action · reasoning (pro/con) — whichever are relevant — plus two scores, always: impact — render Label (value): Minimal (0.25) · Low (0.5) · Medium (1) · High (2) · Massive (3) — and confidence (0.00–1.00).

Step 3 — Synthesize, band, and triage

• Dedup across lenses: the same finding raised by two lenses is one row — keep the max confidence as its confidence c.
• Band each finding by c — the bands are a cost lever, floor at 0.70:

<!-- source: references/confidence-bands.md#mode-f --> - keep (no triage) — c ≥ 0.80. - triage — 0.70 ≤ c < 0.80. - drop — c < 0.70. <!-- /source: references/confidence-bands.md#mode-f -->

• When the triage band has findings, invoke the triage skill via the Skill tool once, passing each banded finding (id = finding #, claim = the finding text) plus the artifact path(s); an empty band loads nothing. Route the returned verdicts:

- consider → walk · skip → park (show in the table, walk only if asked).

• Table (walk set + parked skips):

Verdict is consider/skip for triaged rows, — for the c ≥ 0.80 keep band. Conf. is adjusted_confidence where triage ran, else c.

• Order the walk set by confidence descending.
• Dependency override: if acting on one finding invalidates a later one, walk it first.
• Dropped (c < 0.70): one line each — finding · lens(es) · c.

Step 4 — Walk findings one at a time

For each walk finding (keep band or triage consider, confidence descending; then parked findings if the user asks):

• Present each in this shape:

**Gap [#] — [lens(es)]:**
- Finding: [what's missing]
- Action: [recommended action]
- Impact: [Label (value)] · Conf: [0.xx]
- Triage: [verdict — reason]
- Span: "[quoted artifact text]"

(Omit the Triage line when triage didn't run on this finding; omit Span when the finding isn't location-specific.)

• Decide: AskUserQuestion — apply / defer / drop. Drop later findings an applied decision invalidates, with a one-line reason.
• Check limit: on the first pushback, invoke the second-opinion skill via the Skill tool on the contested action — at most once per finding.
• After the check: re-present apply / defer / drop without another second-opinion call.