engineering-paper-auditor
Audit English engineering manuscripts like a strict reviewer without drafting final prose. Use when the user asks to review, diagnose, critique, find paper weaknesses, check claim-evidence logic, insp
SKILL.mdskills/engineering-paper-auditor/SKILL.md
GitHubEngineering Paper Auditor
Use this skill to find manuscript weaknesses before a reviewer does. It does not produce final paper prose; it produces a ranked list of objections and the cheapest fixes.
Core Stance
- Audit before rewriting.
- Findings must be tied to manuscript text, source evidence, figures/tables, or explicitly missing inputs.
- Do not invent missing experiments, references, mechanisms, line numbers, or completed edits.
- Separate paper-quality blockers from style notes.
- If the user asks whether the manuscript is ready, hand off to
engineering-validationfor build, citation, and final-readiness checks.
Boundaries
- Use
engineering-writingafter audit findings have been turned into section rewrite tasks. - Use
engineering-polishingonly after claim/evidence structure is stable. - Use
engineering-figure-tablefor detailed caption, table, or visual redesign. - Use
engineering-responsefor reviewer/advisor comment replies. - Use
engineering-validationfor readiness, build, reference, and final status.
When to Open Extra Files
| File | Open when |
|---|---|
| ../_shared/reviewer-attack-patterns.md | Any pre-submission review or reviewer simulation |
| ../_shared/experiment-premortem.md | An experiment plan, proposal, or results-table skeleton exists but experiments are not finished |
| references/audit-rubric.md | Running a full manuscript or section audit |
| references/examples.md | Needing concrete audit output examples |
| references/failure-modes.md | Avoiding audit overreach, rewriting, or unsupported conclusions |
| references/story-continuity-audit.md | Auditing whether paper sections form a coherent story |
| references/paragraph-to-paragraph-transition-audit.md | Paragraphs are individually plausible but the section feels jumpy |
| references/claim-resurrection-audit.md | Abstract, Discussion, or Conclusion may revive unsupported claims |
| references/scored-audit-mode.md | The user explicitly asks for scores, a scored review panel, or a scored reviewer simulation |
| ../_shared/story-spine.md | Full-paper or multi-section audit needs story dependency checks |
| ../_shared/evidence-boundary.md | Any claim may exceed available evidence |
| ../_shared/claim-strength.md | Auditing novelty, robustness, causality, or generalization language |
| ../_shared/citation-boundary.md | Related Work, citations, or source support are part of the audit |
| ../_shared/ai-assisted-writing-policy.md | The audit may need to flag AI-smell, disclosure, or author-verification risk |
| ../_shared/list-to-argument.md | A section reads like a list rather than an argument |
| ../_shared/sentence-role-and-story-flow.md | A paragraph may contain redundant, misplaced, missing, or disconnected sentences |
| ../_shared/terminology-ledger.md | Terms, metrics, categories, or method names drift across sections |
Reviewer Simulation (Default For Pre-Submission Review)
Use this whenever the user asks what reviewers will say, whether the paper is convincing, or for a pre-submission check.
- Read the manuscript as a skeptical reviewer from the closest neighboring field. Ask what that reviewer would compare the work against: the simplest alternative configuration, the principled method from the adjacent field, and a baseline given the same capability or prior.
- Walk the patterns in reviewer-attack-patterns.md as prompts. Report any problem the manuscript gives evidence for, whether or not a pattern names it; skip patterns that are not triggered.
- Rank by what real reviewers lead with. In observed review sets, reviewers led with comparison validity (handicapped baselines), reliance on unvalidated assumptions or components, and practical cost (hardware, time, compute). In a small blind test, models without the pattern checklist often put internal-consistency checks (numeric slips, metric definitions) at the top and pushed the reviewers' main objection down to rank 2-7. Rank by consequence instead: a numeric or metric error that changes a headline number, a comparison, or reproducibility belongs with the major objections; only non-substantive slips go to the mechanical list.
- Give the three objections most likely to appear in an actual report first, then at most seven more. Long undifferentiated lists hide the objections that decide the outcome.
- For each objection, give the cheapest pre-submission fix: a matched control, a measurement, an operating-envelope paragraph, a wording change, or a reframing of a weak baseline as a reference configuration.
Reviewer simulation
| Rank | Pattern | Trigger in manuscript | Objection as a reviewer would write it | Cheapest fix before submission |
Mechanical issues: [non-substantive slips: typos, labels that resolve correctly, formatting]
Not triggered: [pattern ids]
Needs author input: [facts the manuscript does not state]
Experiment Plan Pre-mortem (Before Experiments)
Use this when an experiment plan, a proposal, or a results-table skeleton exists but the experiments are not finished. Follow experiment-premortem.md: map each claim to the comparison that would support it, plan matched controls, measurements, logging, and trial counts against P1-P6 and P8, and return the plan table, the cheapest additions, and the claims to drop or narrow now. Choose the mode by what is being reviewed: experiments still to be run get the pre-mortem; arguments and results already written get Reviewer Simulation. A draft with partial results can need both.
Detailed Audit (On Request)
Use when the user asks for a claim-by-claim or section-by-section audit.
- Identify the audit scope: full paper, section, figure/table set, response package, or claim-evidence map.
- Extract the one-sentence thesis and stated contributions.
- Build the claim-evidence and story-spine audits before local style notes.
- Check section jobs and figure/table responsibility against the claims they are asked to support.
- Flag overclaims, missing anchors, section drift, table narration, caption overreach, terminology drift, and unsupported readiness claims.
- Produce a prioritized action list and route each action to the right skill.
Audit verdict
- Overall status: PASS / PASS_WITH_BLOCKERS / CANNOT_DETERMINE
- Highest-risk issue:
- Not verified:
Claim-evidence audit
| Claim | First stated | Method anchor | Experiment anchor | Figure/table anchor | Status | Repair route |
Prioritized action list
| Priority | Action | Owner skill | Required input | Stop condition |
Scored Audit Mode
Off by default. Use it only when the user explicitly asks for scores, a review panel, or scored reviewer simulation. The protocol, lens table, and band anchors are in references/scored-audit-mode.md.
- Three reviewer lenses pass independently: method rigor, experimental evidence, contribution and positioning.
- Bands are heuristic triage anchored to listed findings, not acceptance predictions; no weighted total is produced.
- Material that was not inspected is
CANNOT_DETERMINE, not a low band.