verymath/ai4math-writing/paper-writing

Paper Writing

Use this skill to help write mathematical papers from verified sources, notes, proof artifacts, experiment logs, or reviewer feedback. The workflow is for paper writing; unsupported-claim control is a required review step, not a separate public skill.

Use When

Use this skill when the user wants to:

• draft or revise a mathematical paper section;
• turn proof notes, experiment logs, or reading notes into paper text;
• prepare related work, positioning, abstracts, introductions, or conclusions;
• respond to reviewer comments with evidence-backed changes;
• audit a draft for unsupported claims, missing citations, unclear assumptions,

or overstated contributions.

Do not use this skill to invent theorems, experiments, citations, author claims, or results that are not supported by the provided evidence.

Inputs

• Required: writing goal, target section or deliverable, and source materials.
• Optional: target venue, audience, style constraints, citation policy, reviewer

comments, theorem/proof files, experiment logs, figures, or bibliography.

If source materials are missing, stop and ask for them or produce only an explicit outline of needed evidence. For recurring work, start from templates/source-packet.md.

Outputs

• Draft or revised paper text.
• Claim-evidence notes for substantive mathematical, empirical, or literature

claims.

• A list of unsupported, uncertain, or citation-needed statements.
• Suggested next edits and required source checks.

Subskill Routing

Read skills/registry.yaml when the task needs a specialized audit.

Bundled Templates

• templates/source-packet.md: collect the evidence packet before writing.
• templates/rapid-prototype-paper-skeleton.md: plan the paper structure,

contribution spine, and result dependency map.

• templates/claim-evidence-ledger.md: trace claims to proof, citation,

experiment, or explicit uncertainty.

• templates/proof-obligation-ledger.md: audit assumptions, dependencies,

external result fit, and uncovered cases.

• templates/submission-readiness.md: check build, source package, arXiv, and

venue readiness.

Workflow

1. Identify the writing task: paper skeleton, section draft, revision, related

work, theorem exposition, experiment narrative, or response letter.

1. Inventory the source packet: notes, papers, proofs, code logs, data, figures,

bibliography, and reviewer comments.

1. For early-stage papers, build a rapid skeleton before drafting: contribution

spine, section map, result dependency map, and explicit gaps.

1. Extract only supported claims, definitions, assumptions, results, limitations,

and citation anchors.

1. Load a subskill when the task needs a skeleton, proof-obligation audit,

focused ledger, notation check, formula rewrite, or LaTeX audit.

1. Draft or revise the target text in the requested style.
2. Run a support audit: mark every important claim as supported, needs citation,

needs proof, needs experiment, or should be softened/removed.

1. Return the text plus review notes, not just polished prose.

Review Rules

• Preserve mathematical conditions and quantifiers.
• Distinguish theorem, conjecture, experiment, intuition, and future work.
• Treat "has a citation or proof sketch" as weaker than "all assumptions and

proof obligations are covered".

• Do not fabricate citations, venues, author positions, numeric results, or

proof status.

• Mark uncertain statements explicitly instead of filling gaps.
• Keep AI contribution wording factual and modest when writing acknowledgments,

limitations, or methodology sections.

Failure Modes

• Missing source packet: produce an evidence request, not a full draft.
• Conflicting sources: show the conflict and ask which source is authoritative.
• Unverified proof or experiment: write provisional language and list the

verification still needed.

• Citation gap: mark citation needed with the claim it supports.