cfd-ai-paper-skills

name	cfd-ai-paper-skills
description	Use when writing, reviewing, auditing, revising, or producing LaTeX for CFD-AI and scientific-ML manuscripts, reviews, experiment plans, reproducibility audits, related-work positioning, reviewer responses, or benchmark evaluations.
version	v0.6.5

CFD-AI Paper Skills

Umbrella router for the CFD-AI paper skills package. Use this root skill as the installable entrypoint when the task involves CFD-AI, scientific ML, neural operators, PINNs, surrogate modeling, turbulence closure, flow reconstruction, CFD acceleration, geometry-aware models, flow control, or journal-paper production.

Trigger Conditions

Use this package when the user asks for any of the following:

• write, revise, or review a CFD-AI/SciML manuscript, abstract, introduction, method, result, discussion, or limitations section;
• audit unsupported claims, reviewer risks, reproducibility, experiment design, validation splits, baselines, diagnostics, figures, or LaTeX build readiness;
• plan or critique CFD-AI experiments, benchmark tasks, ablations, held-out regimes, geometry generalization, mesh transfer, solver coupling, or physical diagnostics;
• synthesize related work, build a review-paper taxonomy, position a contribution, or convert a citation dump into a CFD-native evidence map;
• detect and rewrite AI-ish manuscript language so it follows field-native, gold-paper-like CFD-AI/SciML prose patterns;
• produce a LaTeX manuscript seed, response-to-reviewers letter, revision plan, journal checklist, or benchmark review output.

Do not use this package for generic grammar polishing unless the task also needs claim-evidence alignment, scientific accuracy boundaries, CFD/SciML validation logic, or journal-review risk analysis.

Progressive Disclosure

Read only the files needed for the requested artifact first:

1. Start with this router and identify the task type.
2. Read the relevant subskill SKILL.md files from the routing table.
3. Read only the references, rubrics, examples, and templates named by those subskills.
4. Load gold-paper notes only when they directly inform the task's writing pattern, experiment logic, figure logic, or reviewer-defense pattern.
5. Use rubrics and templates after the task route is known, not before.

Unknown scientific facts stay TODO. Do not infer solver settings, mesh details, train/test splits, DOI values, dataset licenses, benchmark numbers, author roles, or corresponding-author status from context.

Routing Table

Task type	Start with	Add when needed
Manuscript writing, abstract/intro rewrite, or AI-ish prose repair	skills/scientific-journal-writing/SKILL.md	references/gold-paper-style-patterns.md, rubrics/gold-paper-closeness-rubric.md, examples/generic-ai-to-gold-paper-prose.md
Unsupported claims, overclaim rewrites, claim-evidence maps	skills/paper-claim-auditor/SKILL.md	skills/cfd-ml-paper-reviewer/SKILL.md, rubrics/claim-evidence-rubric.md, templates/claim-evidence-map.md
Decision-oriented reviewer audit, rejection-risk gate, submission readiness	skills/reviewer-audit-toolkit/SKILL.md	templates/reviewer-audit-report.md, rubrics/reviewer-audit-rubric.md, skills/paper-claim-auditor/SKILL.md, skills/cfd-ml-paper-reviewer/SKILL.md
Reproducibility and numerical-method audit	skills/cfd-reproducibility-checker/SKILL.md	rubrics/cfd-reproducibility-rubric.md, references/field-terminology-style-guide.md
Experiment design	skills/experiment-design-for-sciml/SKILL.md	skills/sciml-experiment-auditor/SKILL.md, templates/experiment-plan.md, rubrics/sciml-experiment-rubric.md
Experiment audit, validation split critique, benchmark-risk review	skills/sciml-experiment-auditor/SKILL.md	skills/cfd-reproducibility-checker/SKILL.md, references/cfd-ai-benchmark-landscape.md
Related-work positioning or citation-map construction	skills/related-work-cartographer/SKILL.md	skills/related-work-synthesis/SKILL.md, references/gold-standard-paper-map.md
Review article or taxonomy synthesis	skills/related-work-synthesis/SKILL.md	skills/scientific-journal-writing/SKILL.md, skills/figure-and-result-storytelling/SKILL.md
Figure/result story, table design, visual evidence logic	skills/figure-and-result-storytelling/SKILL.md	rubrics/figure-evidence-rubric.md, references/figure-archetypes.md
LaTeX manuscript or PDF-oriented production	skills/latex-paper-production/SKILL.md	skills/scientific-journal-writing/SKILL.md, compile/build checks, selected templates
Reviewer simulation and journal-risk report	skills/reviewer-simulator/SKILL.md	skills/cfd-ml-paper-reviewer/SKILL.md, templates/reviewer-report.md
Response to reviewers	skills/response-to-reviewers/SKILL.md	skills/paper-revision-loop-manager/SKILL.md, templates/response-letter.md
Revision planning and manuscript-change management	skills/paper-revision-loop-manager/SKILL.md	skills/reviewer-audit-toolkit/SKILL.md, skills/scientific-journal-writing/SKILL.md, templates/reviewer-editor-loop-report.md, rubrics/reviewer-editor-loop-rubric.md
Multi-agent reviewer-editor workflow	skills/multi-agent-paper-orchestrator/SKILL.md	references/multi-agent-paper-roles.md, templates/multi-agent-paper-workflow-report.md, rubrics/multi-agent-paper-workflow-rubric.md
Ethics, limitations, and LLM disclosure	skills/limitations-ethics-llm-disclosure/SKILL.md	journal instructions supplied by the user
Benchmark review output or package evaluation	evaluation/tasks/*.md matching the topic	evaluation/scorecards/cfd-ai-skill-scorecard.md, relevant subskills above

Output Schema

Choose the contract that matches the route. If the user asks for multiple artifacts, keep each contract separate rather than merging them into prose.

Claim-Evidence Map

Claim	Evidence supplied	Missing evidence/TODO	Reviewer risk	Safer wording	Location

Reviewer-Risk Report

Severity	Issue	Evidence boundary	Likely reviewer objection	Required fix	Suggested rewrite

Reviewer Audit Gate

Gate	Decision	Evidence boundary	Minimum rescue path
Submission readiness	accept / revise / reject-risk
Fatal blockers	count
Major blockers	count

Use templates/reviewer-audit-report.md when the user asks whether a manuscript is submission-ready.

Reproducibility Checklist

Item	Known	Missing/TODO	Why it matters	Manuscript location

Experiment Matrix

Claim tested	Dataset/flow case	Split axis	Baseline	Metric/diagnostic	Ablation/UQ	Failure mode	Reproducibility TODO

LaTeX Manuscript Seed

• complete main.tex source when requested;
• standard packages unless the user supplies a template;
• TODO bibliography entries for unverified citations;
• figure placeholders that state the claim supported by each figure;
• explicit compile notes and no external image dependency unless supplied.

Response Letter

Reviewer point	Response strategy	Evidence needed	Manuscript change	Draft response	Status

Benchmark Review Output

Axis	Evidence expected	Evidence supplied	Failure mode caught	Score/decision	TODO

Gold-Paper Manuscript Seed / Style Reconstruction

Section	Gold-paper move expected	Current draft evidence	AI-ish marker	Required rewrite/TODO

Reviewer-Editor Loop

Stage	Output
Reviewer gate	accept / revise / reject-risk + Fatal/Major blocker count
Editor rewrite	revised abstract/section with unsupported claims removed or bounded
Re-audit	before/after risk and residual blockers
Next actions	experiments/tables/figures needed to rescue remaining claims

Use templates/reviewer-editor-loop-report.md for full reviewer→editor→reviewer cycles.

Multi-Agent Paper Workflow

Agent	Artifact
Reviewer	Fatal/Major rejection risks
Evidence Auditor	claim-evidence map and TODOs
Experiment Planner	validation/baseline/diagnostic rescue plan
Figure/Table Editor	evidence-bearing figure/table sequence
Prose Editor	field-native revised text
Gatekeeper	residual-risk re-audit

Use templates/multi-agent-paper-workflow-report.md when the user asks for explicit multi-agent structure.

Anti-Patterns

• Generic polishing that makes claims smoother without checking evidence.
• Inventing solver settings, benchmark numbers, DOI values, dataset licenses, author roles, citation metadata, hardware/runtime, or train/test split details.
• Treating resource lists, library features, or citation density as proof of benchmark quality.
• Making unbounded claims such as "generalizable", "real-time", "physically consistent", "state-of-the-art", "robust", or "solves turbulence" without same-scope evidence.
• Collapsing different validation axes into one generic "accuracy" claim.
• Treating random snapshot performance as geometry, regime, mesh, BC/IC, time-horizon, or solver-coupled generalization.
• Producing LaTeX/PDF deliverables without a compile/build check or explicit note that the check could not be run.
• Hiding unknowns in polished prose instead of marking them as TODO.
• Letting generic AI academic phrases stand in for field-native problem, gap, evidence, and limitation moves.

Verification Checklist

• Relevant subskills were read before drafting the artifact.
• Reviewer-audit tasks produce a gate decision, Fatal/Major blocker count, and minimum rescue path.
• Reviewer-editor loop tasks produce a reviewer gate, an editor rewrite, and a re-audit with residual blockers.
• Multi-agent paper workflows keep role outputs separate until the orchestrator merge and gatekeeper re-audit.
• Every major claim is mapped to evidence, limitation, or TODO.
• CFD/SciML validation axes are explicit: regime, geometry, mesh, BC/IC, time horizon, QoI, diagnostics, UQ, runtime, and solver coupling where applicable.
• Resource lists, tools, datasets, and benchmark suites are distinguished by source scope and evidence role.
• Reviewer-risk terms are either avoided, bounded, or rewritten.
• Tables/checklists include missing-evidence columns when facts are incomplete.
• LaTeX deliverables include compile/build evidence or a clear reason the check could not be run.
• Output stays within the evidence packet and does not invent scientific facts.
• Manuscript prose can be scored by rubrics/gold-paper-closeness-rubric.md when the task includes writing or style reconstruction.