claude-plugins

Personal Claude Code plugin marketplace by Wenhao Chai.

writing — publication-prep tools

Four skills, all auto-loading by context match.

style — default English-prose standards

Auto-triggers on any writing task the user will send or publish: emails, message drafts, posts, docs, grant proposals, paper prose. 17 canonical English-prose rules apply everywhere. RULE-01 to RULE-12 are distilled from Strunk & White, Orwell, Pinker, and Gopen & Swan, sourced from agent-style under CC BY 4.0; RULE-13 to RULE-17 are added. 20 page-cap additions and 2 final-pass audit rules apply only to a page-capped conference paper.

paper-voice — prose that does not read as LLM-written

For research-paper prose, loaded with style. Fifteen rules measured on the abstract, introduction and first method section of 15 pre-2023 papers by Kaiming He, Joseph E. Gonzalez and Alex Dimakis: authors in the text, consequences as sentences rather than ", so" tails, hedges and a stance, a scene or question before "In this paper", one contribution list, short paragraphs, parentheses for asides. Rules 17-28 distil Alex Dimakis's notes "Ideas for better writing" (2018): a reader simulator, "who cares", one story, self-contained sections, experiments that teach, no conceptual spaghetti, no adverbs, no citation as a noun. Where it disagrees with style, it wins.

python3 writing/skills/paper-voice/voice_stats.py <paper-dir or .tex files>

prints each count beside the corpus and stars the ones far off; --list also prints the sentences to look at first.

preflight — mechanical pre-submission checks

python3 writing/skills/preflight/preflight.py <paper-dir> --venue iclr

Six deterministic checks: latexmk compile gate, page count against the venue cap, build-log scan for undefined citations and overfull boxes, leftover placeholders and \textcolor{red} marks, correspondence-problem density per file, and .bib hygiene. Exit status 1 on any failure, so it chains into a larger gate.

plot — one chart style for posts and papers

style.py and one worked example. The look follows Epoch AI's charts, with type sizes, margins and gaps measured off Epoch's exports: Instrument Sans (bundled, SIL OFL), a light grid in both directions, only the baseline axis, values sitting just above their grid lines, the quantity above each column, panel names inside the panels, and a layout set in inches so every figure has the same rhythm. Every colour is one of Google's GM2 tones, lighter as the inked area grows. Epoch's other forms (subtitles, footnotes, legend columns, export ticks, callouts, direct labels) are options of the same calls. SKILL.md holds the twelve rules, starting with "minimal first": no labels, notes, or reference lines until the owner asks for them.

anti-autoresearch — integrity forensics for paper output

Adversarial self-audit gate: /anti-autoresearch <paper-dir> builds a span-anchored evidence ledger, fans out 11 auditor skills (consistency, citation, baseline, experiment, eval design, proof and derivation, presentation, AI-style impressions, adversarial case, novelty advisory), and hands every finding to a deterministic adjudicator. Loop: draft, audit, fix each verdict-bearing finding, re-run until CLEAN_GIVEN_EVIDENCE. reviewer-discipline is the write-up side: the six CVPR reviewer errors, nine phrasings a review must never contain, the ARR specificity rewrite, and the COLM and TMLR guards.

Vendored from wanshuiyin/Anti-Autoresearch (MIT); tools, references, schemas, and eval live under anti-autoresearch/support/, see its README for the two local path patches.

kexue-fm — offline library of kexue.fm

A SQLite snapshot of every post on 科学空间 (Su Jianlin's blog, 2009 to the last sync): full text as Markdown with LaTeX intact, tags, and the site's official citation. One script and one kexue skill on top:

python3 kexue-fm/scripts/kexue.py search Muon MuP 学习率 --since 2023-01-01
python3 kexue-fm/scripts/kexue.py show 8265
python3 kexue-fm/scripts/kexue.py sync

Each search argument is a separate query; results from all queries merge by reciprocal rank fusion. Skill and docs are in Chinese. Post content is © 苏剑林 under CC BY-NC-SA; the snapshot is for personal study and retrieval.

training-monitor — reading a pretraining run

One text skill, no code. It restates Chunyuan Deng's "Reading a Pretraining Run" (2026) as doctrine an agent can apply to any training stack: the nine first-screen signals, the full metric table with cadence and priority (P0 every step, P1 every 100 steps, P2 on demand), the formulas, the cross-rank reduction rules, a robust spike detector, the fixed triage order, and the common mistakes. The agent inventories what its stack already logs, fills the P0 gaps first, sets thresholds from healthy runs, and decides for itself whether a script is warranted.

presentation — talk decks and explainer videos

Two skills. talk-slides builds, revises and audits a talk deck in the claude.ai Slides artifact format: the author's rules for the slides and the spoken script, layout builders with a template, a static lint, and a render audit in the Slides runtime. video makes a short explainer video of a page from its own animations and data: the author's design rules for motion pieces, a director workflow, and a frame sink that collects the frames for ffmpeg.

python3 presentation/skills/talk-slides/scripts/build_template.py OUT
python3 presentation/skills/talk-slides/scripts/lint.py DECK_DIR
python3 presentation/skills/talk-slides/scripts/audit_render.py RUNTIME_DIR --out shots
python3 presentation/skills/video/scripts/frame_sink.py FRAMES_DIR DIRECTOR_DIR

who-ate-my-flops — PyTorch job performance

Makes one PyTorch training or inference job faster without changing what it computes. /who-ate-my-flops:init pins down the job, the goal and the correctness checks with you and writes them to contract.md. /who-ate-my-flops:diagnose profiles the job, names the one thing it loses the most time to, and proposes a single fix with a measured number. /who-ate-my-flops:optimize repeats profile, edit and correctness check until nothing worth doing is left, then hands back one clean branch. Six supporting skills cover the baseline, torch.profiler analysis, the computation graph, a benchmark script, the final branch, and a PR description that goes to a file and is never pushed. Records land in workspace-who-ate-my-flops/ inside the target repo.

Vendored unchanged from OpenPerfAgent/who-ate-my-flops at 20577a5; the Codex manifest and assets/ are left out.

Install

/plugin marketplace add wenhaochai/claude-plugins
/plugin install writing@wenhaochai
/plugin install anti-autoresearch@wenhaochai
/plugin install kexue-fm@wenhaochai
/plugin install training-monitor@wenhaochai
/plugin install presentation@wenhaochai
/plugin install who-ate-my-flops@wenhaochai

Upgrading from writing 1.x: the 12 integrity-forensics skills moved out of writing into the new anti-autoresearch plugin, so install that one too to keep them. review and the paper-overleaf agent were removed; their mechanical checks became preflight and their venue norms became reviewer-discipline.

License

MIT for plugin code. writing/skills/style/SKILL.md redistributes rules from agent-style under CC BY 4.0, attribution preserved in that file. anti-autoresearch/support/ carries its upstream MIT license in LICENSE.upstream. kexue-fm/data/kexue.sqlite holds third-party content under CC BY-NC-SA. who-ate-my-flops/ keeps its upstream Apache-2.0 license (© 2026 Impossible, Inc.) in who-ate-my-flops/LICENSE.