AI Research Writing Skill

Mandate

Treat paper writing as claim-evidence engineering, not prose generation.

• Ground claims in repository files, experiment artifacts, notes, or verified citations.
• Never invent citations or numeric results. Mark missing evidence visibly and weaken unsupported claims.
• For full papers, produce concrete LaTeX, BibTeX, figure, table, review, and build artifacts; do not stop at an outline.
• Keep human judgment in the loop for scientific correctness, visual inspection, reviewer risk, and final submission decisions.

Route the Task

Choose the smallest route that covers the request. Load only the listed references first.

Execution Contract

For full-paper and submission work:

1. Create paper_state.json from the contract in references/artifacts.md; declare mode, stage, venue, main TeX, bibliography, required artifacts, blockers, and build record.
2. If research_handoff.json exists, run check_research_handoff.py --require-unblocked first. Otherwise inventory the repository and experiments before drafting. Create one task packet and update plan/progress.md at the start and end.
3. Establish the story, claim boundaries, and literature position before long-form prose.
4. Follow the two-pass drafting order in references/workflow.md: write evidence-bearing core sections first, then finalize framing sections.
5. Run the mode-aware quality gate. A terminal stage is invalid when inputs are missing, claims/citations remain unresolved, verification is pending, blockers remain, or build hashes are stale.

Keep process instructions in plan/review files, never in manuscript prose.

Gates

• Evidence: exact numbers must trace to data, logs, tables, notebooks, or scripts. Quantitative papers maintain numeric-evidence v2 selectors and recomputable aggregates.
• Story: define thesis, gap, contribution boundary, and claims to avoid before a full draft.
• Literature: position close work before writing Related Work.
• Citation: scholarly verified records require citation_requests.json and a fresh citation_lock.json; metadata verification and sentence-level support are separate checks.
• Figures: generated images may explain concepts but cannot create numeric evidence. For overview, method, framework, pipeline, architecture, and teaser figures, use built-in image generation as the default paper asset; use deterministic plots/tables for results.
• Review: ground objections in located evidence, separate facts from inference, and state what evidence is needed to assess the claim. For author revision/submission, resolve high-severity objections or record blockers; a completed peer review may recommend rejection.
• Build: compile the real package with record_build.py --run before submission-ready; preserve command, log, tool version, exit code, and hashes.
• Completion: do not say "done" until required artifacts and checks pass or explicit blockers are reported.

Evidence Boundaries

• Treat negative, mixed, missing-seed, or incomplete results as scope constraints, not material to hide.
• Verify that each citation supports the attached sentence, not merely the topic.
• Save important papers locally only when access and redistribution permit; otherwise record stable metadata and access status.
• Generated diagrams communicate a workflow or idea; they are never experimental evidence.

Finish

Report the concrete artifacts changed, checks run, and remaining scientific risks. Preserve the author's intent, but not unsupported wording.