Prepare Conference Manuscripts
Carry one evidence-linked paper through the conference lifecycle. First identify the venue, track, year, stage, and permitted action. Then lock the scientific contract, select one mode, apply current official rules, and verify the exact rendered artifact.
Select one mode
| Mode | Responsibility | Read first |
|---|---|---|
recon | Resolve current official venue rules, template, stages, policy conflicts, and live-form unknowns | manuscript lifecycle |
draft | Turn an author-approved outline and locked scientific contract into manuscript sections when venue policy permits | manuscript lifecycle, abstract and statistics |
audit | Falsify the claim-to-evidence map, venue compliance, anonymity, and artifact consistency without editing | adversarial audit |
polish | Make local evidence-preserving edits to author text without changing the scientific contract | manuscript lifecycle |
rebuttal | Build a complete concern-to-evidence matrix and response or discussion revision | rebuttal and revision |
camera-ready | Reconcile accepted claims, required changes, authorship, disclosure, source, and final PDF under current instructions | manuscript lifecycle |
package | Audit the exact main PDF, supplement, code/data materials, checklist, and archive intended for submission | adversarial audit |
Use one primary mode per pass. Run recon whenever the venue, year, track, stage, or
official authority is uncertain. A venue profile is a dated engineering aid, not a
substitute for current first-party instructions.
For a compound whole-paper request that combines polishing and packaging, whether or not it names audit, follow the whole-paper workflow and return the ordered whole-paper deliverables. It composes existing modes as separate passes, starting with intake and audit before editing, and defaults to a revision copy after the audit boundary.
Resolve the package and venue component
The directory containing this SKILL.md is the installed skill root; call it
$SKILL_DIR. Every bundled component, guide, profile, and command below resolves from
that directory. Do not resolve a tool through a sibling package or a repository parent.
Select a local adapter by exact venue/year/track scope. Read its official-source
record before its profile or proceedings observations.
| Local adapter | Dated scope | Guide | Optional paper observations |
|---|---|---|---|
| AAAI | AAAI-27 Main Technical Track | guide | observations |
| ICLR | ICLR 2027 Conference | guide | observations |
| ACL | ACL 2026 Main Conference through ARR | guide | observations |
| CVPR | CVPR 2026 Main Conference | guide | observations |
| ICML | ICML 2026 Main Track | guide | observations |
| NeurIPS | NeurIPS 2026 Main Track | guide | observations |
| ICCV | ICCV 2025 Main Conference | guide | observations |
If the requested venue/year/track differs from the dated scope, if a profile is stale,
or if its current authority is uncertain, run recon before drafting, auditing, or
packaging. Use generic venue-profile only when
no current first-class profile matches; create a dated task-local profile from current
official sources and do not mutate a bundled profile to impersonate another venue.
The linked paper observations are optional empirical drafting evidence. They can inform writing choices after official authority is resolved, but they never define venue rules, mandatory section order, or a replacement for current first-party instructions.
Record official sources as VERIFIED, UNVERIFIED, COULD_NOT_OPEN, or NOT_FOUND.
Search snippets, prior-year instructions, and remembered limits are not authority.
Preserve conflicts instead of choosing the convenient rule. Do not bundle or modify
venue-owned style files unless redistribution and modification are explicitly allowed.
Freeze the scientific boundary before writing
Obtain the question, contribution, method or argument, data, split, comparison,
independent unit, metric, uncertainty, results, limitations, and supporting artifacts.
Treat them as the locked scientific contract. Manuscript work may explain or narrow
that contract, but it must not create a result, change a protocol, select a favorable
slice, strengthen a causal claim, or convert not_run into completed evidence.
Build a claim map for the title, abstract, contributions, main findings, and conclusion. Each claim needs its exact table, figure, theorem, analysis, or external artifact. Keep load-bearing definitions, protocol choices, and decisive evidence in the main paper. An appendix or supplement is not guaranteed reviewer attention.
For draft, work from paragraph jobs: question, evidence, interpretation, and boundary.
For polish, use targeted edits and a reverse outline. Preserve every number,
comparison direction, uncertainty statement, citation scope, and qualifier. Follow any
narrower venue rule on AI-assisted writing.
For single-section or section-by-section polish, or "continue the next section" / "继续下一节", read section polish and continuation. Keep one manuscript/candidate and the authorized section queue in the existing project checkpoint; verify each changed section before accepting it. Continue through the whole queue when already authorized, then perform one closing cross-section check. Section completion alone is not whole-paper or submission-package completion.
When drafting or repairing a method paper's contribution, Method or Experiments explanation, read method argument cases. Use its inspected paragraph and figure anchors to connect the scientific question, method operation and comparison; do not impose a universal paper template.
Handle abstracts, statistics, and responses
Read abstract and statistics before drafting or auditing an abstract or statistical statement. Use the local abstract component for an abstract checker and the statistics-reporting component for the analysis-register validator. The abstract is a compressed claim map, not an importance pitch. Statistical and evaluation reporting must name the independent unit, estimand, uncertainty, exclusions, and multiplicity treatment actually used. Experimental design, split construction, and frozen evaluation remain upstream in the publication pipeline; this skill reports and audits them without redesigning the protocol.
For rebuttal, read rebuttal and revision. Account for every material
comment. Distinguish correction, clarification, valid new evidence, principled
disagreement, partial resolution, and accepted limitation. Never describe a promised or
requested experiment as completed. Keep the response, revised manuscript, and change
locations synchronized.
Audit source, rendering, and package
Use the local manuscript-core component with a
current venue profile. Its TeX, PDF, archive, and review-report commands live under
$SKILL_DIR/components/manuscript-core/.
Automated checks are necessary but incomplete. Inspect the rendered PDF at the actual
page size and normal reading scale for visible identity, page transitions, fonts,
clipping, equations, tables, figures, captions, links, references, disclosure, and
main-supplement consistency.
Run the adversarial lenses in adversarial audit. A passing format check does not establish scientific soundness, and a strong scientific argument does not waive venue rules. Report observed defects separately from inference. Give each finding the smallest valid action and a verification step.
Preserve the last accepted manuscript before a rebuttal or final revision. For
camera-ready and package, compare the candidate against that baseline and reconcile
all promised changes. Uploading, saving to a submission portal, posting a response, or
submitting is an external action and requires explicit author authorization.
Independent completion and optional specialists
This skill completes conference prose, manuscript-context figures, current venue compliance, responses and the local submission package without another installed skill. For authorized plotting or diagram construction, read standalone figures. New or redesigned schematics default to GPT Image 2.5 concept, then native editable PPTX using available image and presentation tools directly. Produce the master, manuscript exports and inspected render, not merely a brief. Measured plots stay source-derived. Read-only audits do not authorize drawing or changing an approved figure.
When available, build-scientific-visualizations adds specialist layouts and design
helpers to that same default. Its absence does not stop concept generation or local
PPTX construction. Scientific design, new
experiments, protocol changes and disputed result resolution remain separate research
scope, whether handled directly or with optional research-publication-pipeline.
