Tools

Behavior guidelines drawn from Karpathy's observations of how LLMs go wrong when coding: think before acting, keep it simple, make surgical changes only, and drive work by verifiable goals. Can also be used directly as a CLAUDE.md.

Four guidelines based on Karpathy's observations of common LLM coding mistakes: think before coding, prefer simplicity, make only necessary changes, and work toward verifiable goals. Ships as a plugin skill or works as a CLAUDE.md. Community-made, not by Karpathy.

Official Claude Code plugin for the prompts.chat library: search prompts and Agent Skills from your coding assistant, fill in variables and improve prompts with AI. Includes commands, subagents and a remote MCP; saving your own prompts needs an API key.

Makes the agent answer in terse 'caveman' style to cut output tokens, compressing only prose while leaving code, commands and errors intact. Includes a local proxy that compresses logs, test output and other input. Works with Claude Code, Codex and 30+ other agents.

Turns common prompts such as summarizing, extracting insights and rewriting into reusable patterns you pipe text, web pages or YouTube transcripts through, using many model providers or local Ollama. Can also run as a REST API. Run fabric --setup first to configure models.

Plugin for building LLM apps: LangGraph agents with memory and tool use, hybrid-search RAG, vector index tuning, prompt optimization and LLM evaluation. Includes 3 subagents and 8 skills.

CLI and library for evaluating and red-teaming LLM apps: compare prompts and model outputs in bulk, scan for security vulnerabilities and plug into CI/CD, with evals running locally. Now part of OpenAI and still open source under MIT.

Distills a colleague, friend or public figure's chats, documents and public material into a persona profile packaged as a callable skill, for handoffs and preserving know-how. Collects material from Lark, DingTalk, Slack and more. Mind privacy and consent when using others' data.

Pressures the agent with big-tech PUA and PIP-style rhetoric plus structured debugging steps and initiative checks, pushing it to change approach after failures and show evidence before claiming done. Modes include always-on, encouragement and P9 lead. Works with Claude Code, Codex and more.

Context and harness engineering skills covering context compression and optimization, multi-agent architectures, memory systems, tool design and LLM-as-judge evaluation, activated automatically by task. Aimed at building production-grade agent systems.

Writes precise prompts for a specific AI tool: identifies the target tool, extracts your intent, asks at most three questions if details are missing, then outputs a lean, paste-ready prompt. Covers dozens of tools including ChatGPT, Claude Code, Cursor, Midjourney and Sora.

Simon Willison's LLM CLI and Python library: call OpenAI, Claude, Gemini and local models from the terminal, with conversations logged to SQLite automatically, plus embeddings, structured extraction, tool calling and a plugin system.

Connects to the Arize Phoenix observability platform so agents can query LLM traces and sessions, manage prompts, browse datasets and read experiment evaluation results. Maintained by Arize; needs a self-hosted or cloud Phoenix instance.

Rewrites a CLAUDE.md so Claude Code follows it better: keeps project identity and map bare at the top, wraps task-specific rules in narrowly conditioned <important if> blocks, cuts snippets and lint-enforceable rules, and keeps every command.

Uses unified rule files in a .rulesync directory as the source to generate rules, MCP, commands, subagents, skills and hooks configs for dozens of AI coding tools including Claude Code, Cursor, Copilot and Codex. Can also import existing configs or convert directly between tools.

Comet's official Opik MCP: query traces and spans of LLM apps in chat, log scores and comments, and manage datasets, experiments and prompt versions. Works with Opik Cloud (API key or browser sign-in) and self-hosted open-source Opik.