rai-paper-flow

name	rai-paper-flow
description	Routes Robotics & AI conference paper writing for ICRA, RSS, CoRL, ICML, ICLR, and NeurIPS style submissions. Use when the user asks to outline, draft, revise, audit, rebut, package, or prepare a research paper for Robotics or AI venues.
disable-model-invocation	true

RAI Paper Flow

Use this as the user-facing router for paper writing and revision.

Default to the smallest useful pass. Do not run a full paper audit unless the user asks for it, the paper is near submission, or the next blocking issue cannot be found with a lighter pass.

Intake

Before routing, identify:

• mode: quick, standard, or deep; default to quick when unspecified;
• target venue family and page/format constraints if known;
• paper state: idea, remote project, experiment dossier, outline, partial draft, full draft, experiment package, camera-ready package, or review response;
• evidence state: experiments planned, running, complete, or missing;
• risky zones: novelty, related work, baselines, ablations, provenance, citations, figures, coherence, repetition, limitations, submission packaging, or writing clarity.

If the user asks for "make this paper better", first classify the paper state and choose one blocking artifact.

Efficiency Modes

Mode	Use when	Skill budget	Output
quick	Daily writing, section edits, "what should I fix next?", or early draft review.	1-2 skills. Avoid detailed references unless needed.	Top blockers, direct edits, and one next action.
standard	Normal paper revision across structure, evidence, and prose.	2-4 skills. Use references only for selected checks.	Focused audit tables plus repair plan.
deep	Pre-submission, rebuttal-critical, artifact release, or full-paper red team.	As many skills as needed. Load detailed references.	Full audit artifacts and residual-risk table.

When mode is unclear, choose quick and say what a standard or deep pass would add.

Daily Passes

Use these user-facing passes instead of exposing every atomic skill by default.

Pass	Default route	Expand only when
Draft pass	venue-paper-outline / abstract-introduction-builder / related-work-positioning / manuscript-structure-auditor / scientific-writing-editor	The draft lacks shape, positioning, coherence, or readable prose.
Survey pass	evidence-matrix-builder -> survey-synthesis-builder	The source base supports a taxonomy, synthesis claims, and section contracts.
Evidence handoff pass	experiment-dossier-builder -> paper-draft-builder	Results live on a remote machine or a completed dossier needs to become a story and draft.
Claim-evidence pass	citation-integrity-auditor / benchmark-audit / limitations-failure-case-auditor	Claims depend on citations, benchmark results, deployment scope, or limitations.
Reviewer-risk pass	paper-red-team-review	The user wants likely rejection risks or a harsh review.
Submission pass	reviewer-response-builder / latex-submission-checker	Reviews arrived, rebuttal is due, or source package is near submission.

Route by State

Paper state	Route	Do not proceed until
Goal, audience, scope, or deliverable is unresolved	research-plan-grill	The decision brief records the blocking choices and next paper artifact.
No clear contribution	research-plan-grill -> research-idea-rubric	The thesis, novelty claim, baselines, and falsification test are stated.
Code and results exist only in a remote project	experiment-dossier-builder on the remote machine	One portable dossier records the snapshot, method, result ledger, claims, failures, and evidence gaps.
Experiment dossier exists but no coherent draft exists	paper-draft-builder	The thesis, contributions, section plan, and major claims map to dossier evidence or visible TODOs.
Contribution exists but paper shape is unclear	venue-paper-outline	Each contribution maps to evidence and a paper section.
Abstract or introduction is weak, inflated, or missing	abstract-introduction-builder	Problem, gap, approach, evidence, contribution, and scope are explicit.
Related work does not defend novelty	related-work-positioning	Prior-work clusters and claim boundaries are source-grounded.
Experiment claims are central	benchmark-audit	Baselines, metrics, ablations, and result-claim support are audited.
Results must be reproducible or released	experiment-provenance-auditor	Paper numbers trace to code, configs, data, seeds, logs, and plots.
Method, configs, evaluation, or figures may disagree with code	paper-code-consistency-auditor	Material paper-code discrepancies have evidence, severity, and repair paths.
Draft exists with citations	citation-integrity-auditor	High-risk citation issues are marked before polishing.
Draft feels repetitive, disjointed, or hard to follow	manuscript-structure-auditor	The claim thread, repetition map, and structural repair actions are clear.

In quick mode, choose the first row that exposes a blocking issue and stop after the smallest useful artifact. In deep mode, chain rows only when their outputs are needed by the next check.

Venue Stance

Target Robotics & AI conference norms, not Nature-style journal prose by default. Prioritize:

• problem formulation and contribution clarity;
• method reproducibility;
• baselines and ablations;
• benchmark or robot setup validity;
• limitations and failure cases;
• concise related-work positioning.

Paper Gates

• Do not strengthen novelty language until related work and baselines support it.
• Do not hide weak experiments with writing edits; label missing baselines, metrics, ablations, or failure cases.
• Do not summarize benchmark wins before checking baseline fairness and provenance.
• Do not generate citations or bibliographic details from memory.
• Do not accept figures that imply mechanisms, causal arrows, or system behavior not supported by the text.
• Do not sentence-polish text before checking citation support, structure, and repeated claims when those are in scope.
• Do not promise AI-detector evasion; reduce generic, inflated, and formulaic prose while preserving scientific meaning.
• Do not promise rebuttal changes, camera-ready edits, or artifact releases that cannot be verified.
• Do not optimize for one venue's style if the target venue is unknown; state venue assumptions.

Handoff Artifacts

In quick mode, usually do not create files unless the user asks. In standard or deep mode, use or create these artifacts when useful:

paper/
|-- experiment-dossier.md
|-- story-brief.md
|-- paper-draft.md
|-- outline.md
|-- related-work-position.md
|-- abstract-intro-draft.md
|-- claim-evidence-map.md
|-- benchmark-audit.md
|-- provenance-audit.md
|-- paper-code-audit.md
|-- citation-audit.md
|-- structure-audit.md
|-- writing-edit.md
|-- figure-plan.md
|-- limitations-audit.md
|-- red-team-review.md
|-- reviewer-response-plan.md
|-- submission-check.md
`-- talk-storyboard.md

Completion

For quick mode, finish with:

## Top Fixes
1.
2.
3.

## Direct Edits or Repair Actions

## Next Best Pass

For standard or deep mode, finish with a paper status table:

| Area | Status | Blocking issue |
| --- | --- | --- |
| Contribution | | |
| Method | | |
| Experiments | | |
| Provenance | | |
| Figures | | |
| Citations | | |
| Related Work | | |
| Limitations | | |
| Structure | | |
| Writing | | |
| Submission Package | | |

Done means the mode and route are explicit, the next blocking paper artifact is identified, and unsupported claims are not silently rewritten.