Every new Claude Code chat starts with no memory of your code, so Claude searches and opens your files again, one by one. Graphify (★ 141,639 on GitHub, free and open source) reads your project once and builds a map of it, so Claude can follow the map instead of digging. 1. What it does - Type /graphify in Claude Code (or Cursor, Codex, Gemini CLI, Copilot and 20+ more) and it maps your project into a knowledge graph: every file, every function, and how they connect (calls, imports, inherits). - Your code is parsed on your own machine with tree-sitter. Building the code map uses zero AI credits and nothing leaves your computer. - Docs, PDFs, images and video can go in the same map, but that part uses your assistant's model (or an API key you set). 2. Install it (needs Python 3.10+ and uv or pipx) uv tool install graphifyy graphify install The second command finds your assistant and adds the /graphify skill. (No uv? pipx install graphifyy works too.) 3. Build your map Open your project in Claude Code and type: /graphify . You get three files in graphify-out/: - graph.html: open it in a browser and click around your code - GRAPH_REPORT.md: key concepts, surprising connections, questions to ask - graph.json: the full map Claude queries instead of re-reading files 4. Use it Just ask Claude about your code as usual. You can also ask the map directly: - graphify explain "SomeClass": what it is and everything it touches - graphify path "A" "B": how two things connect, hop by hop - graphify query "how does login work?" Keep it fresh: run /graphify . --update after big changes, or graphify hook install to rebuild on every commit. 5. The honest numbers - On ERPNext (about 1M lines), a coding agent with the map found 82% of the key facts vs 71% with plain search and read (6 questions, Graphify's own benchmark). - Pasting the whole repo into every message costs roughly 20x the tokens and finds less (same benchmark). - It is not a guaranteed 70x token cut. Ignore anyone who promises that.

People are editing their videos with Claude now. Here's the whole setup from the reel, step by step, with the three prompts. You need the Claude desktop app and a paid Claude plan (Claude Code isn't on the free plan). 1. Open Claude Code on your computer - Download the Claude desktop app. Not the browser: the browser version can't see the files on your computer. - Open the Code tab, choose Local, then Select folder. 2. Make one folder Create a folder called AI Video Editor with three folders inside: - raw footage - references - exports Select it in Claude Code. Now Claude can see and use your files. 3. Prompt 1 · Setup (installs the five tools) Paste this into Claude Code. You don't install anything yourself; Claude asks before each install. Set up this folder as my video editing workspace. Install and verify these tools on my computer, asking me before each install: - FFmpeg (cutting, trimming, exporting) - Whisper (local speech-to-text, for word-level captions) - yt-dlp (to download my reference videos) - Remotion (React-based video rendering) - HyperFrames (HTML-based motion graphics) Use the package manager that fits my OS (Homebrew on Mac). After each install, run a version check and tell me it works. If something fails, explain why in plain words and fix it. Finish with a short README in this folder listing what's installed and how you'll use each tool. The five tools have 529,981 GitHub stars between them (Nov 2, 2026): - Remotion: builds the video in code (free for individuals and companies of up to 3 people; bigger companies need a licence) - HyperFrames: motion graphics from HTML - FFmpeg: cuts and exports - Whisper: writes the captions - yt-dlp: downloads your reference videos 4. Prompt 2 · Taste (teach it your style) Pick 3–5 videos whose editing style you love and paste their links in. Here are 3–5 videos whose editing style I love: [link 1] [link 2] [link 3] Download them into references/ with yt-dlp. Study each one: hook length, cut pace (cuts per second), caption style (font, size, position, colours, how words highlight), zooms and motion graphics, music energy, and how it ends.

A developer, Jake Schincariol, built a whole trading desk inside Claude: 11 skills, each with one job. It's free and open source (MIT). The data tools need no signup and no API key; only the Order Executor connects to a broker account (Alpaca, paper trading by default). 4,446 stars on GitHub as of Nov 2, 2026. Repo: https://github.com/Jakeschincariol/day-trader-skill 1. Install it - Easiest: paste this into Claude: https://github.com/Jakeschincariol/day-trader-skill install skill - Or in Claude Code, as a plugin: /plugin marketplace add Jakeschincariol/day-trader-skill /plugin install day-trader-skill@day-trader-skill - It needs Python 3.8 or newer, with nothing to pip install. The SEC tools ask for your name and email in trader/config.json (SEC's rule for its free data). 2. The eleven - Chart Scanner: scans about 5,900 US stocks for setups - News Monitor: breaking news on every stock you watch - Entry Timer: tells you the moment a setup triggers - Exit Planner: sets your stop and your target before you buy - Time Machine: replays your trade 10,000 times before you risk a dollar - Risk Manager and Budget Manager: your limits and your position size - Order Executor: paper trading by default, never without your yes - Price Alerts and Earnings Analyst - Pelosi Tracker: reads the stock trades members of Congress have to report (a report can arrive up to 45 days after the trade) 3. The Tesla test from the reel The repo's own example is a real Tesla setup from Oct 19, 2026: a breakout above the morning range, entry 364.18, stop 358.32, target 370.04. The Time Machine replayed the rest of that day 10,000 times, built from Tesla's own last 59 sessions: - target first: 29% - stop first: 25% - neither by the close: 45% Its own verdict: close to a coin flip. I rebuilt the Time Machine from the repo's documented method and re-ran it on the repo's Tesla data. Same numbers: 29.38 / 25.48 / 45.14.

REA (github.com/morluto/rea) connects your AI agent (Claude Code, Codex, Cursor, Gemini CLI and others) to reverse-engineering tools. You point it at an app you're allowed to analyze, ask how a feature works, and it shows you the evidence. It's free (MIT licence), it runs on your own laptop, and it had 411,931 stars on GitHub when I checked on 31 Oct. Set it up (5 minutes) Install Node.js (version 22.23.4+, 24.22.1+ or 26.12.1+). Open a terminal in a normal folder, not inside the app you'll analyze, and run: npx rea-agents setup Pick your agent, read the changes it proposes, approve them, and restart your agent. Only get REA from github.com/morluto/rea, rea.tools or the "rea-agents" npm package. Look-alike repos exist. My test, so you can repeat it • The app: Chrome's offline dinosaur game, the open-source build (github.com/wayou/t-rex-runner, BSD licence, from Chromium). I compiled and minified it first, so my agent saw no source code. • The question I asked: "In this game the screen sometimes turns dark. When exactly does that happen, how long does it last, and how does the code make it happen?" • The result, 4 min 54 s: it found the exact code. When your score hits a milestone, the whole page flips to night (a moon and stars fade in), and a timer flips it back. Then it wrote its own working copy of the feature. • What it didn't find: the actual numbers. Those are in the file: every 700 points, 12 seconds of night. Ground rules Use it on your own software, open-source software, or anything you have permission to analyze. Many app licences forbid reverse engineering. Native apps (not JavaScript) need an analysis engine such as Ghidra or Hopper. REA's setup can install Hopper if you approve it. Questions? Drop them below 👇