ML Paper Studio
An evidence-based writing skill group for computer science and ML/AI papers. One entry point coordinates six stages, from research code and results to a Chinese argument structure, literature positioning, an English manuscript, and a checked LaTeX/PDF deliverable.
Version: 1.1.0 · Skill: $ml-paper-studio · License: MIT
What it does
| Stage | Work | Main outputs |
|---|---|---|
| 0 · Venue | Check the official requirements for the exact venue, year, track, paper type, and submission phase; obtain the template. | Requirements, source links, template provenance. |
| 1 · Evidence and outline | Inspect the supplied code, protocols, and actual results; distinguish observed, planned, synthetic, and incomplete evidence. | Chinese outline, claim–evidence map, terminology, figure plan, gaps. |
| 2 · Literature and positioning | Research prior, closest, and competing work; check citation metadata and the evidence behind claims. | Search record, source and comparison matrices, defensible contribution positioning. |
| 3 · Chinese draft and review | Draft from the confirmed framework, question the argument, simulate review, and make concrete revisions. | Chinese manuscript and review/change record. |
| 4 · Academic English | Translate and polish the manuscript, abstract, introduction, and captions; adapt to the venue. | English manuscript and checks for changes to numbers, equations, citations, and claim strength. |
| 5 · LaTeX and PDF | Build an editable LaTeX project, export the actual PDF, and check its content, references, layout, and submission requirements. | Source project, PDF when build/export tools are available, and a check report. |
For the full workflow, the default author decision point is after Stage 2: review the revised Chinese outline, contributions, closest-work comparison, and evidence boundaries. An existing confirmation is reused; explicit delegation to the assistant is also supported. After confirmation, an authorized full workflow continues without asking to start every stage.
Self-review defaults to two rounds—evidence/argument review, then simulated peer review—with at most a third round for specific remaining issues. Simulated review is not formal peer review. You can also request a single stage or section without restarting the whole workflow.
Citation checks distinguish three questions: Does the work exist? Was the relevant text retrieved? Does that text support this particular claim? A working DOI alone does not answer all three. Literature positioning compares components, their coupling, constraints, and end-to-end evidence; “a complete system” alone is not a novelty argument.
Writing while experiments continue
The optional parallel drafting mode develops the full paper structure from the author's declared design while remaining experiments run. Keep work-progress notes in an internal ledger; the manuscript can use mature academic prose without repeatedly saying that experiments are unfinished. Unknown empirical findings retain linked result placeholders rather than invented numbers or conclusions. An abstract deadline can be served by a result-neutral abstract, subject to the actual venue rules; an internal full-abstract draft may keep result slots. Draft-stage checks use --purpose draft; existing strict checks remain unchanged and submission readiness is not assessed by the draft check.
Install
Recommended: ask Codex to install the skill
In Codex with $skill-installer available, send:
Use $skill-installer to install the skill from
https://github.com/polarskicpl/ml-paper-studio/tree/main/skills/ml-paper-studio
Use it on the next turn after installation. To install manually, copy the complete skills/ml-paper-studio/ directory to a skill directory supported by your client. Current Codex documentation uses ~/.agents/skills/ for personal skills and .agents/skills/ for project skills. Some older clients and the bundled installer still use ~/.codex/skills/ (or $CODEX_HOME/skills/). Follow your client's supported location; preserve existing local changes and avoid duplicate installations. See the official skills guide.
Optional: install through a plugin marketplace
This repository also includes a portable plugin.json and .agents/plugins/marketplace.json. In a supported Codex client, register the repository as a marketplace source:
codex plugin marketplace add polarskicpl/ml-paper-studio
Then select this marketplace in the client's plugin directory and install the plugin. Registering the source does not install it. Start a new session after installation. This is a self-hosted marketplace source, not a listing in an official curated directory; client compatibility has not been verified across all clients. See the official plugin guide.
Choose either the direct skill installation or the plugin route to avoid loading duplicate copies.
Use
Start with your research materials:
Use $ml-paper-studio. Inspect this project's actual code, frozen protocols,
and existing results. Produce a Chinese paper outline, a claim–evidence map,
and a list of gaps. Do not start new experiments. I will specify the venue later.
Or request one stage:
Use the literature stage of $ml-paper-studio. Find prior and closest work for
this design, verify the references, and compare component-level ideas with
the overall mechanism. Return the positioning and comparison matrix in English.
Do not draft the full paper or run experiments.
使用 $ml-paper-studio，只将这份中文方法与实验章节改成学术英文。
保留数值、公式、引用所支持的句子和结论强度，不重跑其他阶段。
The core stage instructions are written in Chinese. You can request English outputs; the default full workflow develops the Chinese framework and draft before English revision. More examples: usage guide.
Requirements and validation
- Bundled helper scripts use the Python standard library. Python 3.10+ is recommended; the current checks ran on Python 3.14.4. Additional tools may be needed for the manuscript itself.
- Literature and venue verification require accessible sources. PDF delivery requires an available TeX build/export tool and PDF inspection/rendering tools. The workflow is not tied to a particular desktop tool; missing checks remain explicitly unrun.
- Validation includes 63 regression cases for the workspace-state helper and a synthetic-material outline trial. Other helper checks retain their documented limits. A real-paper, end-to-end PDF workflow has not yet been validated. Structural checks do not establish scientific validity or submission compliance.
See validation scope and the reproducible check command.
The skill does not authorize new training runs, changes to frozen experiments, or publication of private research materials. Follow the project's own rules and the user's actual authorization.
Repository and attribution
- Main skill: routing and workflow boundaries.
skills/ml-paper-studio/stages/: the six internal stage modules.skills/ml-paper-studio/references/: artifact contracts, review rubric, examples, and source research.skills/ml-paper-studio/scripts/: workspace checks, template fetching, translation drift checks, and static TeX auditing.
The original files in this repository are released under the MIT License. The workflow was informed by research into 16 skill directories across seven repositories; the source notes record what was adopted or omitted. Upstream skill snapshots, scripts, and conference templates are not bundled; links do not change their respective licenses.