Academic paper

Template-native paper workflow. It uses project manuscript sources, generated variables, validation gates, and evidence registries rather than a copied external academic agent prompt.

Natural invoke

• "Help me plan this paper"
• "Write the outline and methods section from these artifacts"
• "Revise the manuscript after reviewer comments"
• "Check citations and convert the final paper format"
• "Write an AI-use disclosure tied to the repo evidence"

Inputs to confirm

• Project - resolve through docs/_generated/active_projects.md.
• Mode - plan, outline, full, revision, format, citation-check, prose-quality, or disclosure.
• Manuscript style - Pandoc-crossref or registry tokens; do not mix styles.
• Evidence sources - generated variables, claim ledger, figures, tables, references, validation reports.

Workflow

1. Plan first when uncertain - capture target venue, paper type, contribution, section map, and unresolved material gaps.
2. Draft from artifacts - write against projects/<project>/manuscript/, generated variables, figure captions, and citekeys; use [MATERIAL GAP] instead of fabricating missing evidence.
3. Preserve injected values - prefer [[VAR:...]], registry-backed labels, citekeys, and project-defined cross-reference tokens over hard-coded numbers.
4. Validate continuously - run prerender, markdown, citation, PDF, and integrity checks appropriate to the mode.
5. Self-edit the prose - run the AI-writing-fingerprint scan (prose-quality) and tighten any flagged stock phrasing, em-dash over-use, or uniform sentence length. Treat it as advisory (revise, do not blindly delete); it is a deterministic clean-room distillation of the ARS writing-quality check, not a model judgment.
6. Verify references exist - in citation-check/format/finalize, run the reference-existence gate so no cited DOI/arXiv id is fabricated and no metadata is mismatched.
7. Handoff for review - use academic-paper-reviewer for read-only critique; use claim verification before finalization.

Deliverables

• Plan mode: section outline, evidence map, unresolved gaps, and commands to validate.
• Full/revision mode: source-layer manuscript edits, response notes, and validation status.
• Format/citation mode: citation audit, render-safe changes, and output paths.
• Disclosure mode: concise AI-use disclosure grounded in project evidence and human review checkpoints.

Verification commands

bash
uv run python -m infrastructure.validation.cli prerender projects/<project>/manuscript --repo-root .
uv run python -m infrastructure.reference.citation validate projects/<project>/manuscript/references.bib
uv run python -m infrastructure.reference.verification verify projects/<project>/manuscript/references.bib --live --as-of-year <year>
uv run python -m infrastructure.validation.cli prose-quality projects/<project>/manuscript
uv run python -m infrastructure.validation.cli markdown projects/<project>/manuscript --repo-root . --strict
uv run python -m infrastructure.validation.cli integrity output/<project>/

References

• MODE_REGISTRY.md
• manuscript-creation
• manuscript-claim-verification