scientific-writing-skills

Scientific Writing Skills

A Claude plugin for scientific writing — six AI skills that edit research prose, structure IMRaD manuscripts, run the writing process, handle peer review and publication ethics, draft grant proposals, and translate research for general audiences.

Distilled from all eight units of Stanford's Writing in the Sciences (Kristin Sainani) — 62 lecture transcripts, ~141,000 words — into instructions an AI assistant can actually act on, plus four Python scripts that catch mechanically what a reviewer would catch by eye.

Works with Claude Code, Claude Cowork, and any tool that supports the Claude Agent Skills format.

What it looks like

A paragraph from examples/, before and after sciwrite-prose:

Before

As it is well known, the implementation of a hardware acceleration strategy provides a reduction in the inference latency of convolutional neural networks that are deployed on resource-constrained embedded platforms. However, a degradation of the detection accuracy is caused by the quantisation of the network weights to eight-bit integers. Therefore, there is a significant increase in memory pressure when the batch size is raised above four. Furthermore, the results were not inconsistent with those reported for comparable architectures, and the overhead was not significant. Moreover, the proposed pipeline for real-time defect detection, which combines a lightweight backbone with a hardware-accelerated non-maximum suppression stage running on the programmable logic of the Kria KV260, achieves 31 frames per second at 1080p. The DPU IP was integrated into the PL via AXI4, and the PS ran a PetaLinux image with the XRT stack managing DMA transfers between DDR and BRAM. Due to the fact that a majority of the successful solutions rely on this approach, we adopted it.

After

Hardware acceleration cuts inference latency for convolutional neural networks on resource-constrained embedded platforms, but quantising the weights to eight-bit integers costs detection accuracy, and memory pressure climbs sharply once the batch size passes four.

Our pipeline pairs a lightweight backbone with hardware-accelerated non-maximum suppression on the programmable logic of the Kria KV260, reaching 31 frames per second at 1080p. The deep-learning processing unit talks to the programmable logic over AXI4, and the processing system runs a PetaLinux image whose runtime moves data between DDR and block RAM. These results agree with those reported for comparable architectures, and the overhead stays small. We chose this design because most published solutions rely on it.

Same claims, same numbers, a third shorter. The skill named every one of those 25 defects and said why; the script found them without reading for meaning. It also added no number the author never gave: "the overhead was not significant" stays vague in the rewrite, and the skill asks the author for the figure instead of inventing one.

What it does

You paste a paragraph and ask why it feels bloated. It tells you: a smothered verb, a throat-clearing opener, a nominalisation, and a buried predicate — and hands you the rewrite.

You ask where to start on a paper. It tells you the tables and figures come first, and why.

You get a decision letter that looks like a rejection. It tells you that "revise and resubmit" is the outcome you were aiming for, and drafts the point-by-point response.

you   → "review the discussion section of my thesis chapter"
claude → reads the .tex, extracts that section, runs the audits,
         reports what to fix in order of impact

The six skills

Each skill keeps its SKILL.md short and loads detailed reference files only when the task needs them, so asking for a one-paragraph cleanup does not drag eight units of doctrine into context.

The scripts

Four Python scripts (standard library only, no dependencies) do the mechanical detection so the model spends its attention on judgment:

Everything supports --json. They report candidates, not verdicts — passive voice is legitimate in Methods, a repeated keyword is mandatory, and PSNR may be standard in your venue.

python scripts/prose_audit.py paper.tex --section discussion
python scripts/keyword_consistency.py paper.tex --terms "obese group,control arm"
python scripts/numeric_consistency.py paper.tex --json

Install

Any agent (Claude Code, Cursor, Codex, Copilot, Gemini CLI and others)

npx skills add dvdsosa/scientific-writing-skills

The skills CLI asks which agents and which skills to install; add -g to install for your user rather than for the current project, or --skill sciwrite-prose to pick one skill.

Claude Code plugin

claude plugin marketplace add dvdsosa/scientific-writing-skills
claude plugin install sciwrite

Or from a local clone:

git clone https://github.com/dvdsosa/scientific-writing-skills.git
claude plugin install ./scientific-writing-skills

Running the scripts

The scripts live inside the sciwrite-prose skill (skills/sciwrite-prose/scripts/, also reachable as scripts/ at the repository root). The skills find them wherever sciwrite-prose was installed — plugin, npx skills add for a project or globally, or a manual copy into ~/.claude/skills/. Install sciwrite-prose alongside any other skill you choose; without it, sciwrite-manuscript and sciwrite-process fall back to doing the checks by reading.

To call the scripts yourself, point at the clone:

python ~/path/to/scientific-writing-skills/scripts/prose_audit.py paper.tex

Pass the master .tex file, not a single chapter: keyword_consistency.py and numeric_consistency.py compare across sections, and an isolated fragment gives them nothing to cross-check.

By hand

Copy the skills/sciwrite-*/ directories you want into ~/.claude/skills/ (or your agent's skills directory). Copy sciwrite-prose too if you want the scripts.

Verify

python scripts/prose_audit.py --help

Python 3.8+. No third-party packages.

Design notes

Traceability. Every rule cites the course module it comes from, so you can check any of them against the source. Conventions of your own — your supervisor's, your group's, your target journal's — are stated in the conversation, and the skills follow them and tell you which course rule they override.

The em dash stays. Some scientific-writing prompts ban it as a marker of machine-generated text. This plugin does not, because the course teaches the dash as the most versatile mark available and demonstrates case by case where commas lose the emphasis and parentheses bury essential information. The real constraint — use it sparingly, reserve it for the hard jobs — is already in the source.

Two example registers. The course's own examples are biomedical and are quoted as such. A second set in skills/sciwrite-prose/references/engineering-examples.md applies the same rules to embedded systems, FPGA design and computer vision, clearly marked as written for this plugin rather than drawn from the course.

Language. The skills are written in English; they answer in whatever language you write to them in.

What this is not

• Not a substitute for reading the course. It is free, it is excellent, and eleven hours of Kristin Sainani will teach you things a plugin cannot.
• Not a content reviewer. It improves how claims are delivered, never the claims themselves. If something looks scientifically wrong it flags it and leaves it alone.
• Not a plagiarism checker. For overlap against your own bibliography, see tools built for that.
• Not affiliated with Stanford University or Kristin Sainani. This is an independent, non-commercial study aid built from publicly available course materials. All pedagogical credit belongs to the course; any errors in transcription or interpretation are mine.

Contributing

Corrections against the source material are especially welcome — if a rule here misrepresents what the course teaches, that is a bug. See CONTRIBUTING.md.

Good first contributions: additional worked before/after pairs from your own field; false positives in the audit scripts; translations of the reference files.

Credits

Course: Writing in the Sciences, Kristin Sainani, Stanford University.

Books the course recommends, and which stand behind much of this material: William Zinsser, On Writing Well; Strunk & White, The Elements of Style; Mimi Zeiger, Essentials of Writing Biomedical Research Papers; Thomas Annesley's series in Clinical Chemistry.

Grant-writing material in sciwrite-proposals derives from guest lectures by Crystal Botham and Sky Brubaker (Stanford Biosciences Grant Writing Academy); science-communication material in sciwrite-outreach from guest lectures by Amy Adams (Stanford Science Communication).

License

MIT for the plugin code and the written instructions. The underlying course is © Stanford University; this repository quotes short excerpts for educational commentary and does not redistribute course materials. See NOTICE.

Keywords: scientific writing · academic writing · manuscript editing · research paper writing · IMRaD · peer review · grant writing · specific aims · lay summary · science communication · Claude skills · Claude Code plugin · AI writing assistant for researchers · thesis writing · journal submission