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).
Plugin who-ate-my-flops from marketplace wenhaochai in wenhaochai/claude-plugins. Commands read from the repo's manifest — 6 days ago.
clean-upDeliver a finished optimization run as a branch the user can merge. Cherry-pick the commits that changed the workload onto the base, hardcode the A/B flags, re-verify on the shipped code, strip the checkpoint calls in a last commit, and leave the user standing on that branch with a handover. Use whe
convert-to-benchmark-scriptTurn the script a user really runs into a benchmark script an optimization loop can run over and over. Same shape as the real job, a few minutes long, repeatable, and leaving nothing behind. It produces the script and a table of what changed, and it does not run it. Use when init or optimize needs a
diagnoseProfile the job defined in contract.md, name the one thing it loses most time to, and propose a single fix with a measured number behind it. It stops at the first finding. optimize is this same work repeated until nothing worth doing is left.
extract-computation-graphRead a PyTorch job's code into two files in the workspace’s records/ directory. records/computation_graph.json says what is computed: the module tree down to op level, with a code anchor and a repeat count on every entry. records/execution_schedule.json says when and where it is computed and what mo
initWork with the user to pin down which PyTorch job gets optimized, write contract.md, and put the correctness instrumentation in place. The optimize skill later reads the contract.md written here and optimizes on its own. Every piece of alignment with the user happens in init, all of it.
optimizeOptimize the torch job defined in contract.md with no further input. Profile, edit, and check correctness in a loop, then hand back one clean optimized branch.
publish-prWrite the pull request description and one issue for a finished campaign into proposed-pr-description.md and proposed-issue.md, for the user to review and post. Says what the PR may claim, what goes in each section and in what order, and how it is worded. Fills template/pr-description.md and templat
setup-baselineBuild the starting point both diagnose and optimize measure everything else against: a worktree, a workspace, the computation graph and execution schedule, and the base commit measured for speed, memory and correctness and profiled in full. It arms nothing, so the caller decides what happens next. U
torch-profileProfile a PyTorch job, training or inference, with torch.profiler. Analyse the profile, read the surface symptoms, chase each root cause down to a file:line, and hand back a diagnosis. This skill produces no patch. Use when a PyTorch job is slow (low MFU, slow steps, idle GPU) and the user wants to
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
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
Harness-native ECC operator layer - 71 agents, 314 skills, 98 legacy command shims, reusable hooks, rules, selective install profiles, and production-ready workflows for Claude Code, Codex, OpenCode, Cursor, and related agent harnesses
200 skills · 98 commands · 71 agents · 1 MCP server · hooks
agentic-flow (external: maintained in ruvnet/agentic-flow, pinned to commit d19737b, plugin 3.4.0): the agentic-flow MCP server pinned to [email protected] (ReasoningBank learning, ecosystem tools prompt_injection_scan, swarm_partition_tasks, sublinear_solve, sublinear_pagerank and ecosystem_status, multi-provider routing, autopilot learning; readonly profile by default, no tool pre-approved), 2 agents, 4 skills, /agentic-flow-status, and a mod guard on agentic_flow_agent with /agentic-flow-guard
4 skills · 1 command · 2 agents · 1 MCP server
Reliable automation, in-depth debugging, and performance analysis in Chrome using Chrome DevTools and Puppeteer
8 skills · 1 MCP server
Perform AI code review on Git diffs — supports workspace changes, branch ranges, and single commits with concurrent per-file analysis, codebase search, and deep context-aware review.
2 commands
Transform Claude into an Azure expert. This plugin integrates the Azure MCP server and specialized Azure skills to move beyond generic advice. It enables Claude to perform real-world tasks: listing resources, validating deployments, diagnosing infrastructure issues, and optimizing costs across 50+ Azure services.
44 skills · 1 MCP server
Systems programming with Rust, Go, C, and C++ for performance-critical and low-level development
3 skills · 1 command · 4 agents
This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.