rai-research-flow

name	rai-research-flow
description	Routes Robotics & AI research work across search, reading, evidence mapping, ideation, paper writing, figure planning, and audits. Use when the user asks to enter a field, build a literature base, brainstorm a project, write a survey or paper, or coordinate multiple research skills.
disable-model-invocation	true

RAI Research Flow

Use this as the user-facing router for Robotics & AI research work.

Intake

Before routing, identify:

• research goal: onboarding, survey, ideation, experiment design, writing, figure work, or audit;
• domain and constraints: Robotics subfield, AI method family, hardware/simulator, timeline, venue, and available sources;
• evidence state: no sources, source list only, reading cards, matrix, draft, figures, or code/results;
• required artifact: onboarding map, search log, reading cards, matrix, idea cards, experiment dossier, outline, paper draft, figure plan, provenance audit, or review report.

If the user asks for "research help" without a concrete artifact, propose the smallest useful next artifact instead of starting a broad open-ended workflow.

Route by Intent

User intent	Route	Gate
Goal, scope, or deliverable remains ambiguous after inspecting available materials	research-plan-grill	A decision brief names the objective, constraints, evidence standard, and next artifact.
Inherit a folder, notes, paper list, or lab knowledge base	knowledge-base-onboarding	Source inventory and trust labels exist before synthesis.
Enter a new area	paper-search-protocol -> paper-reading-card -> evidence-matrix-builder	At least one reproducible search log before synthesis.
Build a survey	evidence-matrix-builder -> survey-synthesis-builder	Reading cards and traceable matrix rows exist before taxonomy or prose synthesis.
Brainstorm a project	research-plan-grill when key decisions are unresolved, then research-idea-rubric	Either an evidence matrix exists or the output must be labeled speculative.
Write a conference paper	rai-paper-flow	Contributions, evidence, and venue assumptions must be stated.
Position related work	related-work-positioning	Prior-work clusters and novelty boundaries use verified sources.
Audit experiments or results	benchmark-audit -> experiment-provenance-auditor	Claims map to result evidence and traceable runs.
Hand off a remote project for local writing	experiment-dossier-builder	The dossier records the inspected snapshot, result provenance, and missing evidence.
Turn a completed evidence handoff into a draft	paper-draft-builder	The story thesis and every major claim map to dossier results or verified literature.
Plan or audit figures	scientific-figure-director	Each figure needs a one-sentence scientific claim.
Build a lab, conference, journal-club, or project talk	slide-talk-builder	Audience, duration, takeaway, source package, and requested deck format are explicit.
Check whether a manuscript describes the implementation accurately	paper-code-consistency-auditor	Paper and code versions are identifiable and material discrepancies have repair owners.
Review a draft	rai-paper-flow in quick mode, then expand to citation or red-team checks only if needed	Do not style-polish unsupported claims.

Phase Gates

• Discover before map: do not build taxonomy from a thin or unlogged search.
• Decide before execute: do not start a costly search, experiment, or draft while its blocking scope or evaluation decisions remain implicit.
• Inventory before onboarding claims: do not summarize a project folder or source set until source trust labels are explicit.
• Map before ideate: do not recommend projects from an unmapped field unless the user explicitly accepts speculative brainstorming.
• Evidence before writing: do not draft survey conclusions until the evidence matrix exists or missing evidence is stated.
• Support before polish: do not polish claims that lack support; mark them as unsupported and ask for evidence.
• Trace before claim: do not strengthen benchmark claims until result provenance and baseline fairness are checked.
• Contract before figure: do not treat a figure as valid until its scientific claim, entities, arrows, and labels pass review.

Handoff Rules

• When switching skills, pass the current artifact paths or pasted content, not only the user's broad goal.
• Preserve uncertainty. If a source, result, or citation is unverified, carry that status into the next skill.
• Prefer one clear next artifact over multiple partial artifacts.
• Ask the user before starting a large search, full-paper review, or multi-skill survey if the scope is not bounded.

Default Artifacts

Use these names unless the repo already has a convention:

research/
|-- onboarding-map.md
|-- search-log.md
|-- reading-cards/
|-- evidence-matrix.csv
|-- survey-synthesis.md
|-- idea-cards.md
|-- benchmark-audit.md
|-- provenance-audit.md
|-- experiment-dossier.md
|-- story-brief.md
|-- paper-draft.md
|-- figure-plan.md
|-- paper-code-audit.md
|-- talk-storyboard.md
`-- audit-report.md

Completion

Finish with:

## Research Flow Status
| Phase | Artifact | Status | Blocker |
| --- | --- | --- | --- |

## Subskills Used

## Next Best Step

Done means the chosen subskills were routed in a defensible order, every produced artifact is named, and any blocked gate is explicit.