wenhaochai

Integrity forensics for research papers, and the reviewer discipline for writing up what it finds. /anti-autoresearch builds a span-anchored evidence ledger, fans out 11 auditors, and hands every finding to a deterministic adjudicator: the model proposes, the adjudicator decides. Detect-only and observability-aware. Vendored from wanshuiyin/Anti-Autoresearch (MIT); support files under support/.

13 skills

Diagnose a PyTorch job on one node and make it faster without changing what it computes: profile, propose a fix with a measured number behind it, or run a whole campaign that measures, gates on correctness and hands back a branch. Vendored from OpenPerfAgent/who-ate-my-flops (Apache-2.0).

9 skills

Publication-prep tools for academic and public-facing writing. Four skills: style (17 canonical English-prose rules, 19 page-cap additions, 2 audit-time rules), paper-voice (the prose voice of pre-2023 papers by Kaiming He, Joseph E. Gonzalez and Alex Dimakis, with a counter that compares a draft to them), preflight (deterministic pre-submission checks on a LaTeX tree), and plot (one matplotlib style for posts and papers after Epoch AI's charts: Instrument Sans at Epoch's measured sizes, light grid, baseline axis only, Google GM2 tones by mark size; style.py plus one example).

5 skills

Talks and videos. talk-slides builds, revises and audits talk decks as claude.ai Slides artifacts, with the author's rules for slides and script, layout builders, a lint and a render audit; video makes a short explainer video from a page's own animations and data, with the author's design rules and a frame sink.

2 skills

科学空间 kexue.fm（苏剑林博客）全站离线库：一个 SQLite（全文 Markdown、LaTeX 原样、标签、官方引用格式）+ 一个 kexue skill（多路检索 search、读全文 show、增量 sync）。

1 skill

Monitoring doctrine for pretraining runs, distilled from Chunyuan Deng's "Reading a Pretraining Run" (2026): nine first-screen signals, a 39-row metric table with cadence and priority, formulas, cross-rank reduction rules, a robust spike detector, the triage order, the common mistakes, and the training re-evaluation curve for reading which data a finished run fits worse. Text only; the agent decides how to implement it for its own stack.

1 skill