# Critical Second Pass > Use when the user explicitly asks for a second, harder look at something already produced: a prior answer, proposal, plan, findings, workflow or governance decision, diff, completed change, handoff, or written draft. Triggers include "re-review", "review again", "pressure-test this", "poke holes", "red-team this", "stress test", "sanity check this", "audit", "what risks am I missing", "can this be improved", "did we miss anything", "再 review 一遍", "二次 review", "再看一遍能不能更好", "挑挑刺", "看看有没有风险/遗漏", and "帮我 pressure test 一下". Do not use for casual first-pass feedback such as "what do you think?" or "review this" without an explicit second-pass or pressure-test intent. - Skill: `flyingfsr/critical-second-pass` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add flyingfsr/critical-second-pass` - Raw SKILL.md: https://api.skillmd.com/api/skills/flyingfsr/critical-second-pass/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - Author: FlyingFSR (https://skillmd.com/u/flyingfsr) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/flyingfsr/critical-second-pass --- # Critical Second Pass ## Goal Re-examine a prior output as if a different agent produced it, and report honestly what, if anything, still needs to change. A second pass is valuable because the first pass is usually written under momentum and mild confirmation bias. The job is to recover the outside view that momentum buried. Do not manufacture findings. `No material changes needed` is a valid and useful result. ## Trigger Boundary Use this only when the user explicitly asks for a second look, re-review, audit, pressure test, risk check, improvement check, or whether a prior output can be improved. Do not trigger on casual first-pass feedback, ordinary clarification, or a generic request to review something for the first time. If the intent is unclear, ask one short clarifying question. ## What To Review If the user names the artifact, review exactly that. If they do not, default to the most recent substantive output in the conversation: the last proposal, finding set, plan, draft, diff, handoff, or completed change. If several artifacts are plausible, name the one you are reviewing in the first line so the user can redirect cheaply. ## Modes Pick the mode that matches the artifact. Some artifacts need both. **Ops / Decision Mode** Use for proposals, plans, workflow or governance decisions, diffs, code, handoffs, and completed implementation work. Test whether the next step will execute correctly and whether the reasoning holds up. **Content / Creative Mode** Use for scripts, articles, posts, essays, and other material meant to be read or heard. Test whether it is true, whether it lands, and whether it sounds like the user rather than generic AI writing. ## Priority Rank by consequence, not cleverness. - **P1** - would change the decision or outcome if left in. - **P2** - worth fixing because it reduces risk, confusion, or future rework, but the work survives without it. - **P3** - optional polish. - **Keep** - working parts that should stay. For content work, P1 includes false claims, off-voice material that reads as AI, an opening that does not earn attention, or a privacy/boundary issue. For ops work, P1 includes stale status, a broken handoff, misleading verification, wrong ownership, or a recommendation that would cause incorrect execution. ## Output Open with exactly one verdict, then the findings: - `No material changes needed` - `Small refinements` - `Needs correction before proceeding` Then provide findings by priority, followed by the single smallest useful next action. Be concise and concrete. Include file references when reviewing files. If nothing material needs to change, say so clearly and name any residual risk instead of inventing work. ## Workflow 1. Identify the artifact under review. 2. Inspect current files, diffs, commits, drafts, tests, or logs when they are involved. Do not review from memory if the artifact can be checked. 3. Stay in scope. Raise adjacent issues only when they directly affect the decision at hand. 4. Give the verdict, findings by priority, and the smallest useful next action. 5. Do not implement changes unless the user explicitly asks you to apply or fix them. 6. Do not recursively review your own second pass unless the user asks for another round. ## Common Risk Checks Use only the checks that fit the artifact. Ops / decision: - A one-off judgment promoted into a global rule. - Too much process, documentation, or record-keeping added. - Stale status, stale handoff instructions, or contradictory source-of-truth surfaces. - A recommendation with no owner, file, action, or decision gate. - Verification weaker than the confidence claimed. - A non-problem solved, or the right problem solved at the wrong layer. - A technically true finding that is not worth acting on. - A fix that risks undoing user-owned or unrelated work. Content / creative: - The agent's framing presented as the user's lived experience. - An emotional beat or claim that rings false. - Generic phrasing that reads as AI. - An opening that does not earn attention. - A privacy or identity boundary crossed without need. ## Convergence On repeat passes, if what remains is only P3 or mostly restates earlier points, say so plainly. Name the residual risk, recommend stopping if that is the right call, and hand the decision back to the user.