Paper Suite
English | 中文
Academic writing, publication-figure, and presentation skills for Codex and Claude Code. Built bottom-up from a 96-paper corpus (CVPR / NeurIPS / Nature / IEEE / classics), with reusable Python scripts and contract-first workflows so agents can produce reliable outputs instead of only giving advice.
What is this
A monorepo of 17 sub-skills that cover the full academic pipeline: writing, figure-making, and presentation. Rules and templates are distilled from real academic papers and packaged references, so the suite can run without a separate corpus checkout. Users can also provide example papers or PDFs for dynamic venue calibration: Paper Suite extracts writing style, palette, layout, and figure conventions before adapting the next draft, figure, or deck.
Three skills, four scenarios each:
| Skill | Sub-skills | What it does |
|---|
| writing | style-k, prose-polish, venue-fit, venue-adapt, abstract-craft, reviewer-eyes, deai-taste, rebuttal-helper | Polish prose, remove AI-tell wording, reshape to venue style, learn style from example PDFs, write abstracts, pre-review manuscripts, respond to reviewers |
| figure | plot-mpl, diagram-tikz, diagram-drawio, venue-switch, figure-anti-defect | Make contract-first matplotlib / TikZ / draw.io figures, switch venue style, auto-check defects |
| ppt | deck-outline, deck-page, speaker-notes, deck-deck | Build editable PPTX decks for group meetings, conference talks, thesis defenses, and funding pitches |
Invocation keywords
Use the prefixed public names to avoid collisions with other installed skills:
| Public keyword | Use it for |
|---|
$paper-suite | Top-level router across writing, figures, and PPT |
$paper-writing | Prose polish, AI-tell cleanup, venue fit, venue-adapt from example PDFs, abstracts, reviews, rebuttals |
$paper-figure | Contract-first plots, Nature/NMI/IEEE/CVPR figure style, TikZ/draw.io diagrams, defect checks |
$paper-ppt | Editable academic PPTX decks, slide outlines, single slides, and speaker notes |
How it differs from the things you may have tried
| Common failure mode | This suite's answer |
|---|
| "Polished" academic writing that reads like a marketing blog | 12 universal rules + 140-word AI-tell blacklist, derived from cross-group statistics on 96 real papers |
| "Looks like SOTA" claims with no numbers | Rule 2: numbers, not adjectives. Rule 9: no "novel". Rule 12: no empty future work. |
| Figures that look like they came out of a Jupyter notebook | 5 matplotlib templates + 4 venue-specific font/width configs, all tested against 96 real figures |
| Figures that drift away from user intent | Contract-first workflow asks the agent to confirm data semantics, encodings, panels, and venue constraints before plotting |
| PPT slides that are 80% bullet points | Editable PPTX templates with academic layouts, timing notes, speaker notes, and PowerPoint compatibility tests |
| Rebuttal letters that promise things you didn't do | 5 comment types + 5 action types, every response cites manuscript section/line |
| Style transfer that loses the actual claim | venue-fit reads 6 venue rule docs and re-shapes; venue-adapt can learn from user examples/PDFs while preserving claims |
The data this is built on
- 96 real papers in 6 groups (CS-vision 20, CS-other 20, classic 10, Nature 13, IEEE 14, cross-disciplinary 9)
- Distilled rules from the corpus are vendored into
references/ files inside each skill
- Optional external corpus support for users who want dynamic venue adaptation
- 2,609 full-text page images (one PNG per page, for visual reference)
- 6-group statistics (median pages / sentences / references / markers, computed with supplementary excluded)
- Total raw data: 1.5 GB
Repository layout
