CCF Paper Writer

Invocation Controls

CCFA Handoff Mode: PARTIAL (Recommended). Follow metadata.ccf_skill_controls.handoff_question_mode, ../ccf-common/references/handoff-modes.md, and ../ccf-common/references/task-modes.md. User scope and existing authorization govern execution; preserve plan-only, no-browsing, exact-output, and no-new-files requests.

This skill owns manuscript text, compression, and presentation prose. Preserve the supplied problem, method mechanism, experimental setting, numbers, citation keys, and conclusion unless a research change is authorized. A requested revision authorizes ordinary accurate wording edits and necessary local checks.

Run ccf-humanization as the first manuscript-facing preflight. Read ../ccf-humanization/references/humanization-policy.md and references/prose-quality-guardrails.md. These guide direct scientific writing; they do not start a separate review or a confirmation loop.

Writing Standard

Write from the contribution: what problem matters, why the mechanism addresses it, what the evidence establishes, and what follows. Replace reviewer-facing reassurance, apologetic novelty claims, repeated caveats, denial-led positioning, and stacked hedges with their scientific payload. Delete sentences with no payload. Do not add a limitations sentence to each paragraph, abstract, caption, or conclusion by habit.

Keep meaningful uncertainty, known negative results, necessary assumptions, fair-comparison constraints, and mandatory disclosures in direct academic prose. Do not write that a method is confirmed, approved, or publication-ready. Explain the actual method and relevant configuration. A missing result stays missing; a proposed study is not a completed experiment. Never guess citations or insert invented measurements to complete a draft.

Use the humanization policy as the single source for punctuation and pattern thresholds. Natural writing is a semantic editing task; passing a phrase checker does not establish quality. Keep scientifically necessary terms, negation, and hedging even when a heuristic flags them.

Modes And Reference Loading

An abstract, caption, synopsis, paragraph edit, or exact JSON response does not require venue guides, full exemplars, reviewer panels, or length planning by default. Load matched exemplar cards from references/exemplars/index.md only for a requested style adaptation or a full manuscript that benefits from them. Never copy exemplar wording or technical content.

Workflow

1. Identify the requested mode, output format, evidence, and target length from the conversation and supplied files. Infer routine choices; ask only for a missing decision that changes the research claim, deliverable, or feasibility. Continue independent work while that decision is pending.
2. For an authorized edit to an existing manuscript, revise that file in place and preserve Markdown/LaTeX sections, commands, citations, labels, equations, and float environments unless restructuring is requested. Follow ../ccf-common/references/artifact-contracts.md; create no dated or v2 copies for an ordinary iteration. Supplying a file for inspection alone does not authorize rewriting it.
3. For a full submission manuscript, read references/venue-guides/index.md and the matched guide, then establish a section budget using references/length-budget-policy.md. If no venue or guide is available, use the existing NeurIPS guide/template as a disclosed drafting assumption. references/output-style-policy.md controls ambiguous format choices. Current final policy requires official verification; local guides can support a provisional draft.
4. Organize the scientific argument. For a substantial introduction or full manuscript, use references/storyline-blueprint.md; for bounded prose, directly connect problem, insight, mechanism, evidence, and implication. Keep material scope where it changes interpretation. Do not generate a reviewer-risk register or multi-expert tournament for ordinary writing.
5. When literature is needed, load references/citation-workflow.md, obtain verified entries through public-safe retrieval, and update the existing bibliography without changing unrelated keys. Use ccf-literature-searcher for concrete retrieval needs within the request. For supplied-evidence-only or no-browsing tasks, preserve provided citations and mark unresolved references without inventing entries.
6. Draft the requested artifact. Use references/research-writing-patterns.md, the relevant references/section-modules.md passages, and applicable references/writing-checklists.md checks for substantial writing only. A full paper includes all scientifically relevant and venue-required sections, with explicit TBD placeholders for unavailable evidence. Keep methods, setups, and analyses substantive; never pad with generic caution, assumed modules, or imaginary results.
7. For reported experimental comparisons, consult ../ccf-humanization/references/experiment-discipline.md when method identity or protocol matters. Use supplied design specifications for a proposed method and cited evidence for prior work; do not demand runnable checkpoints or full experimental confirmation to perform a text edit.
8. For a full LaTeX manuscript with an available engine, compile and inspect page count, errors, references, and affected layouts. Fix concrete issues, expand actual explanatory gaps, and compress excess content. Put auxiliary files, build logs, and current previews under the established build directory or chosen task working directory; export the requested PDF to its canonical path only after a successful build. Preserve existing relative includes and build configuration. Recompile after relevant changes; if another pass makes no substantive progress, revise the approach or report the exact remaining issue. Do not loop to fill pages or chase harmless warnings. Leave final venue compliance to ccf-submission-checker when requested or necessary.
9. Check the resulting prose once with Humanization's sentence decisions. For a full section or paper, run scripts/check_prose_quality.py when available. Inspect defensive-language candidates in context; fix real problems and retain justified scientific language. Use --strict only when a strict lint report is requested, not as a universal publication gate. Rerun only for changed content or unresolved findings.
10. Finish the requested deliverable. If missing evidence requires a new research decision, state the affected claim and continue supported work. Use the existing owner for requested experiment design, visuals, scientific review, or response prose; do not start optional downstream deliverables merely because they appear in a checklist.

Output Contract

Return the actual revised text, manuscript, compressed passage, or presentation prose first in the user's format. Ordinary local edits need no mode/status/next-skill block. For full manuscript files, report their paths, substantive changes, relevant validation, and concrete unresolved evidence. Exact schemas stay exact.

A complete draft contains the requested scientific sections at appropriate depth; it may remain evidence-incomplete. Distinguish draft completion from submission readiness. Do not label unknown results as observations or claim measured score improvements from a prose revision.

References

Load selectively by the modes above:

• references/output-style-policy.md, references/length-budget-policy.md: source-format preservation and manuscript budgeting.
• references/venue-guides/index.md, references/exemplars/index.md: target venue and selected style moves.
• references/storyline-blueprint.md, references/section-modules.md, references/research-writing-patterns.md: scientific organization and section writing.
• references/citation-workflow.md: verified citation insertion.
• references/prose-quality-guardrails.md, references/writing-checklists.md, scripts/check_prose_quality.py: semantic prose review and supporting diagnostics.
• references/compression-rules.md: evidence-preserving compression.
• references/table-style-guide.md: existing LaTeX table source guidance; visual redesign belongs to ccf-visual-composer.
• references/score-lifting-loop.md, references/expert-review-loop.md: applying actual reviewer deductions when that work is requested.
• ../ccf-common/references/review-output-standards.md: conditional scores and frozen review standards, only for requested review-related outputs.