Open-Source AI Repos & MCP Servers

Claude, asked to refactor a loop: 'I'd be happy to help you refactor that function. Let me analyze the code carefully and walk you through what I'm doing step by step...'Claude with Caveman installed: 'Refactor: extract loop to function. Done.'Caveman is a Claude Code skill that cuts token costs by up to 75% by forcing the model to respond in stripped-down, direct language. No articles (a, an, the). No filler words. No pleasantries. No hedging. Created by Julius Brussee. 5,000+ GitHub stars. Hit #1 trending on GitHub. Ranked #1 on Hacker News with 883 points in early April 2026. MIT licensed.The Core InsightPhrases like 'I'd be happy to help you with that' and 'Let me summarize what I just did' contribute nothing to output quality. They burn tokens. They slow responses. They push users into Claude Pro usage limits faster. They are not polite — they are padding.Caveman strips the padding. It injects a compact system prompt rule at the start of every Claude conversation that enforces the caveman style. Output becomes terse, direct, and measurably cheaper to generate.The Four Grunt LevelsCaveman ships four intensity levels, from light to extreme brevity:Grunt 1 (Light): Removes obvious filler ('I'd be happy to...', 'Let me walk you through...'). Keeps articles and complete sentences. Typical savings: 22%.Grunt 2 (Medium): Strips articles where grammatically non-essential. Replaces multi-word phrases with single words. 'In order to' becomes 'to'. Typical savings: 45%.Grunt 3 (Heavy): Eliminates most prose. Responses become bullet-like. 'The function refactors the loop by extracting...' becomes 'Refactor: extract loop.' Typical savings: 65%.Grunt 4 (Extreme): Near-telegraphic output. 'Fixed: bug line 47' instead of full sentences. Typical savings: 75-87%. Not for customer-facing outputs — designed for developer workflows where you just want the answer.Built-In UtilitiesCaveman includes tools that apply the same philosophy to specific workflows. Terse commit messages (no 'chore:', no 'feat:', just the action). One-line PR reviews that identify the issue without restating the code. Input compression that rewrites your prompts to remove your own filler before sending.The input compression is a second win. Even if you do not shorten Claude's output, shortening your own prompts cuts input token costs. Combined with Grunt 3 output constraints, you hit the 65-75% savings range routinely.Backed by ResearchA March 2026 academic paper titled 'Brevity Constraints Reverse Performance Hierarchies in Language Models' found that brief-constraint outputs improved accuracy by 26 percentage points on certain reasoning benchmarks. The theory: verbose outputs encourage the model to generate filler that introduces factual drift. Forced brevity keeps the model on the essential reasoning path.Caveman operationalizes this research. The grunt levels are calibrated against the paper's findings. Grunt 2 tracks the research's 'optimal brevity' level for accuracy gains.InstallationOne line for Claude Code: /skill install juliusbrussee/caveman. Manual install: download the skill file from the GitHub repo and place it in ~/.claude/skills/. Cursor, Windsurf, and Copilot have equivalent install paths documented in the README.Configuration is minimal. Set your default grunt level in the skill config. Override per-conversation with a /grunt 3 command. That is it.Token Savings ExampleA typical coding task: 'Refactor this 40-line function to use a reducer pattern.' Claude default: 850 output tokens explaining the refactor, showing the diff, noting edge cases, suggesting tests. Claude with Grunt 3: 220 output tokens. Pure diff plus a one-line summary. Same correctness. Same usefulness for a developer who knows what they are doing. 74% fewer tokens.At Claude Opus 4.7 pricing ($5/$25 per million tokens), that is a meaningful reduction on a monthly scale. A developer generating 5M output tokens per month drops from $125 to $32 in Claude API costs.When Not to Use CavemanCustomer-facing documentation. Claude's verbose style works for users who want explanation, not answers. Code walkthroughs for learning. Teaching someone new to programming requires the filler that Caveman removes. Legal or compliance outputs where completeness matters more than brevity.Caveman is a developer tool. Turn it on for your coding sessions, turn it off when you need Claude to write human-facing text.LimitationsGrunt 4 output is hard to read for anyone but the developer who wrote the prompt. Sharing Grunt 4 outputs with a team without context causes confusion.The skill is Claude-specific in its tuning. It works on Cursor and Copilot, but the token savings are less dramatic because those editors have their own system prompts that partially override the skill.Caveman reduces output tokens, not latency. Responses are faster because there is less to generate, but the model still processes the full input. For input token savings, use the bundled input compression utility.For more Claude Code skills, browse repos.skila.ai/skills. For AI coding tools that pair well with Caveman, check tools.skila.ai. For articles on token economics and LLM cost optimization, visit news.skila.ai.

43,000 GitHub stars. Trending #1 on GitHub. For a single Markdown file.Andrej Karpathy Skills distills observations from the former Tesla AI director and OpenAI founding member into one CLAUDE.md file that changes how Claude Code writes software. The repo, created by developer Forrest Chang, codifies four principles that address the most common ways LLMs fail at coding tasks.The core insight, in Karpathy's words: "LLMs are exceptionally good at looping until they meet specific goals. Don't tell it what to do — give it success criteria and watch it go."The Four Principles1. Think Before Coding. LLMs make assumptions without checking. They see a function name and guess what it does instead of reading the implementation. This principle forces Claude to state assumptions explicitly, present multiple interpretations when instructions are ambiguous, push back when the request seems wrong, and name confusion before proceeding. The result: fewer "helpful" changes that break things.2. Simplicity First. Left unchecked, LLMs overengineer. They add abstraction layers, create utility functions nobody asked for, and build error handling for scenarios that will never occur. This principle enforces minimum code that solves the actual problem. No speculative features. No unnecessary abstractions. No premature optimization.3. Surgical Changes. LLMs see a file and start "improving" everything. Rename variables. Refactor functions. Add docstrings. None of it was requested. This principle restricts Claude to touching only what the task requires. Match existing code style. Do not improve adjacent code. Remove only code that your changes made unused.4. Goal-Driven Execution. Instead of telling Claude what to do step by step, convert tasks into verifiable success criteria. "Make the tests pass" is more effective than "fix the bug in line 47." Claude can loop, retry, and explore alternatives when it has a clear success condition. It cannot do that with imperative instructions.InstallationTwo methods. The plugin marketplace: /plugin marketplace add forrestchang/andrej-karpathy-skills then /plugin install andrej-karpathy-skills@karpathy-skills. Or the manual approach: download the CLAUDE.md file with curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md and place it in your project root.That is it. No dependencies. No configuration. No runtime. One file changes the behavior.Why 43,000 Stars for a Markdown FileBecause it works. The principles address the four most frustrating LLM coding behaviors: assumptions without verification, unnecessary complexity, unrelated changes, and vague execution. Anyone who has used Claude Code, Cursor, or Copilot has experienced all four. One file, four rules, measurably better output.The 43K stars also reflect a shift in how developers think about AI coding tools. The bottleneck is not model capability. Claude Code is already brilliant at writing code. The bottleneck is behavior — getting the model to write the right code for the right reasons. Karpathy's principles are behavioral constraints, not capability enhancements.LimitationsThe file is optimized for Claude Code specifically. Other agents (Cursor, Copilot, Codex) can read Markdown skill files, but the behavioral changes may be less effective because they have different system prompt structures. The principles are also opinionated — "Simplicity First" conflicts with teams that value defensive programming and extensive error handling. The file does not address multi-file refactoring, test generation, or architecture decisions.For more Claude Code skills, browse repos.skila.ai/skills. For AI coding tools, check tools.skila.ai. For articles on AI-powered development workflows, visit news.skila.ai.

ARIS turns your AI coding agent into an ML researcher that works while you sleep. Literally. Start a research loop before bed. Wake up to a paper draft with scores improving from 5/10 to 7.5/10 across four autonomous review rounds overnight.31 composable Markdown-only skills cover the full research lifecycle. Literature survey scans papers and builds citation graphs. Idea brainstorming generates 8-12 concepts per session. Novelty filtering checks each idea against existing work. GPU pilot experiments run verification on real hardware. Cross-model adversarial review sends your draft to a competing AI for criticism. LaTeX generation compiles the paper. Rebuttal drafting responds to reviewer objections.The cross-model adversarial review is the standout feature. Claude Code writes and executes the research. GPT-5.4 xhigh reviews it. This is not the same model talking to itself. These are two different AI architectures with different blind spots. Claude is fast and fluid. GPT-5.4 xhigh is slower but more rigorous in critique. The adversarial tension between speed and rigor produces better papers than either model alone.How It WorksEach skill is a plain Markdown file. No framework. No dependencies. No lock-in. Copy the skill files into your agent's skills directory and they work immediately. ARIS supports Claude Code (primary), Codex, OpenClaw, Cursor, Trae, and any LLM agent that reads Markdown skill files.Four effort levels control how deep the research goes. Lite uses 0.4x the default token budget for quick passes. Balanced is the default. Max runs at 2.5x for thorough reviews. Beast mode hits 5-8x for maximum-depth research sprints.Research WikiThe Research Wiki maintains persistent memory across sessions. Every paper read, every idea generated, every experiment attempted, and every failed approach gets logged. This creates an anti-repetition knowledge base. The agent does not re-explore dead ends or re-propose rejected ideas. Three skills (/research-lit, /idea-creator, /result-to-claim) hook directly into the wiki for context-aware generation.Safety GatesThree safety mechanisms prevent hallucinated citations. ARIS fetches real BibTeX entries from DBLP and CrossRef before including any reference. If a citation cannot be verified against these databases, it is flagged rather than fabricated. This addresses one of the biggest risks in AI-generated academic writing.LimitationsCross-model review requires API access to both Claude and GPT-5.4 xhigh, which means paying for two providersGPU pilot experiments assume access to GPU hardware or cloud compute; the skill does not provision resourcesv0.3.11 is early. Some skills advertised in the README may have incomplete implementationsResearch Wiki grows without automatic pruning, potentially degrading retrieval quality over timeFor more Claude Code skills, browse repos.skila.ai/skills. For AI research tools, check tools.skila.ai. For articles on AI-powered development workflows, visit news.skila.ai.