double-check
Targeted re-verification of a specific claim, output, or decision. Re-read sources, challenge, cross-reference, and deliver a verdict. Trigger when the user runs /double-check or asks to "verify this", "re-check this", "are you sure about this", "go over this again", or "make sur
Install
npx skills add https://github.com/allemaar/open-skills/tree/main/skills/double-check
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install allemaar-open-skills@llmmart
git clone https://github.com/allemaar/open-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole allemaar/open-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
/double-check
Multi-pass re-verification of a plan, output, decision, or document — re-read from scratch, challenge assumptions, propose improvements, and iterate until solid.
Re-Read & Improve Protocol — not a single-pass gate. The agent re-reads the target fresh, challenges each part, proposes improvements, and iterates. Each pass tightens the result.
Structured execution spec:
protocol.yon. It carries the five phases, the scope/verdict gates, and the rules as validatable records; this file is the explanation. Keep the two in sync — edit one, update the other and refresh the@STAMPdate.
Caller Options. Before executing, run the Caller Options protocol (
caller-options/SKILL.md): triage this invocation for material optionality across the venues and modes declared in front-matter; if one path clearly dominates, proceed silently; otherwise surface the options to the caller. A resolved-invocation marker means COP already ran — execute the fixed combination directly, do not re-enter COP.
Phase 1: Scope
Identify what is being double-checked — a plan, an output, a decision, a document. Load it fresh. Do not rely on prior context or cached understanding.
Gate: Cannot double-check without a clear target. If the target is ambiguous, ask before proceeding.
Phase 2: Fresh Re-Read
Re-read the target from scratch as if seeing it for the first time. Go back to the actual sources — files, code, docs. Do not verify from memory or prior analysis alone.
Summarize:
- What it says
- What it assumes (implicit and explicit)
- What it depends on
List every assumption, stated or unstated.
Phase 3: Challenge
Challenge each part of the target. Be adversarial — actively try to find problems.
For plans: Are any steps wrong, missing, out of order, or unnecessary? For outputs: Are any claims unsupported? For decisions: Were alternatives considered?
Look for:
- Gaps and omissions
- Wrong or unstated assumptions
- Over-engineering or under-engineering
- Scope issues
- Ordering issues
- Missing dependencies
Phase 4: Improve
Propose concrete improvements based on the challenge findings.
For each issue found:
- State the problem
- Propose the fix
- Explain why it's better
If the target is already solid, say so with evidence — but don't stop looking just because it seems fine.
Phase 5: Verdict
Deliver one of three verdicts:
- SOLID — target holds up under scrutiny; cite evidence
- IMPROVED — issues found and fixes proposed; list them; offer to apply and run another pass
- RETHINK — fundamental problems; explain what needs to change
Gate: Double-check is incomplete without a verdict delivered to the user.
Rules
- MUST go back to actual sources when re-reading — never verify from memory or prior analysis alone
- MUST be adversarial when challenging — actively look for problems, don't just confirm
- MUST cite evidence for every conclusion in the verdict
- MUST NOT rubber-stamp without substantive analysis
- SHOULD offer to apply fixes and run another pass when verdict is IMPROVED
Next Steps
/plan-execute— if verdict is SOLID/plan-create— if verdict is RETHINK (redesign needed)
Human output. This skill's handler-facing output obeys the human-output contract (
human-output/SKILL.md).
Self-improvement. On completion, run the Self-Improvement Protocol (
self-improve/SKILL.md): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.
Files (open-skills)
-
protocol.yon 4.8 KB · in bundle
-
SKILL.md 4.8 KB
--- name: double-check description: Targeted re-verification of a specific claim, output, or decision. Re-read sources, challenge, cross-reference, and deliver a verdict. Trigger when the user runs /double-check or asks to "verify this", "re-check this", "are you sure about this", "go over this again", or "make sure this is right". Not /verify (formal phase gate needing intent+plan+execution artifacts) or /insight-critique (advisory review of an output) — double-check is a quick targeted re-read of one claim. caller-options: venue: [inline, delegated] default-policy: ask visibility: public self-improvable: true triggers: - "/double-check" - "verify this" - "re-check this" - "are you sure about this" - "go over this again" - "make sure this is right" next-skills: - skill: verify phrase: "/verify" why: "Escalate from a single-claim re-check to the formal intent/plan/execution gate before proceeding" - skill: github-sync phrase: "/github-sync" why: "If the claim held up, commit and push the verified work" - skill: insight-adversarial phrase: "/insight-adversarial" why: "If the re-check raised doubt, stress-test the output from multiple adversarial POVs" --- # /double-check Multi-pass re-verification of a plan, output, decision, or document — re-read from scratch, challenge assumptions, propose improvements, and iterate until solid. > **Re-Read & Improve Protocol — not a single-pass gate.** The agent re-reads the target fresh, challenges each part, proposes improvements, and iterates. Each pass tightens the result. > **Structured execution spec:** [`protocol.yon`](protocol.yon). It carries the five phases, the scope/verdict gates, and the rules as validatable records; this file is the explanation. Keep the two in sync — edit one, update the other and refresh the `@STAMP` date. > **Caller Options.** Before executing, run the Caller Options protocol (`caller-options/SKILL.md`): triage this invocation for material optionality across the venues and modes declared in front-matter; if one path clearly dominates, proceed silently; otherwise surface the options to the caller. A resolved-invocation marker means COP already ran — execute the fixed combination directly, do not re-enter COP. --- ## Phase 1: Scope Identify what is being double-checked — a plan, an output, a decision, a document. Load it fresh. Do not rely on prior context or cached understanding. **Gate:** Cannot double-check without a clear target. If the target is ambiguous, ask before proceeding. --- ## Phase 2: Fresh Re-Read Re-read the target from scratch as if seeing it for the first time. Go back to the **actual sources** — files, code, docs. Do not verify from memory or prior analysis alone. Summarize: - What it says - What it assumes (implicit and explicit) - What it depends on List every assumption, stated or unstated. --- ## Phase 3: Challenge Challenge each part of the target. Be **adversarial** — actively try to find problems. For **plans:** Are any steps wrong, missing, out of order, or unnecessary? For **outputs:** Are any claims unsupported? For **decisions:** Were alternatives considered? Look for: - Gaps and omissions - Wrong or unstated assumptions - Over-engineering or under-engineering - Scope issues - Ordering issues - Missing dependencies --- ## Phase 4: Improve Propose concrete improvements based on the challenge findings. For each issue found: 1. State the problem 2. Propose the fix 3. Explain why it's better If the target is already solid, say so with evidence — but don't stop looking just because it seems fine. --- ## Phase 5: Verdict Deliver one of three verdicts: - **SOLID** — target holds up under scrutiny; cite evidence - **IMPROVED** — issues found and fixes proposed; list them; offer to apply and run another pass - **RETHINK** — fundamental problems; explain what needs to change **Gate:** Double-check is incomplete without a verdict delivered to the user. --- ## Rules - MUST go back to actual sources when re-reading — never verify from memory or prior analysis alone - MUST be adversarial when challenging — actively look for problems, don't just confirm - MUST cite evidence for every conclusion in the verdict - MUST NOT rubber-stamp without substantive analysis - SHOULD offer to apply fixes and run another pass when verdict is IMPROVED --- ## Next Steps - `/plan-execute` — if verdict is SOLID - `/plan-create` — if verdict is RETHINK (redesign needed) > **Human output.** This skill's handler-facing output obeys the human-output > contract (`human-output/SKILL.md`). > **Self-improvement.** On completion, run the Self-Improvement Protocol (`self-improve/SKILL.md`): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.