# Beginner: your first week with AI on your research **What rstack is.** Claude itself is a harness. It's the thing around the model. But you, for your own work, end up building your own harness, which is the way you like to work. The harness is just text files, Python scripts, shell scripts and hooks, and you build it over time. rstack is mine. You will not trust the AI from day one. So you have to build the trust in it, and that's why you have to start small. At first, you don't trust it. It does stuff bad, and then it starts doing stuff better. That's when you get comfortable giving it more stuff, because your trust meter goes up. If you cannot build a system of trust, then you will never scale. So I built a bunch of safeguards and guardrails around how I work: a lot of scripts, a lot of instructions, a lot of gates. A gate is a yes or no for a lot of stuff that I do. I have a gate for emails that tries to stop the email from going if it doesn't sound like me, and one for analysis: if it didn't do all the steps, it stops it from going out and redoes it. They're not perfect. But you teach the AI how to catch itself and how to do things the way you like it. And I have a very low activation energy to give it a task, and also a low barrier to just throw everything away and start over. You don't need most of this repo yet. If you do research, start with Claude Science. It's really easy, and it's a great first step for your research. The first step is the hardest; after that it's fun. ![Three tracks: beginner, intermediate, advanced, and you move up by what you can do](../../assets/onboarding-tracks.png) ## 1. Try Claude Science with one of your papers [Claude Science](https://support.claude.com/en/articles/16563838-get-started-with-claude-science) is Anthropic's desktop app for research (macOS and Linux, on a paid Claude plan). Install it, start a project, and give it two things: one of your own papers, and one question about it you never had time for. A check you always meant to run, an extension you sketched in the discussion, a dataset you wanted to try it on. Then read what it did, ask it to show the code, and push back where it's wrong. That's the whole first step. If it gives you something useful, you're ready for step 2. ## 2. Install a coding agent When you want an agent working in your own files and code, install Claude Code with one command (macOS or Linux), then check it runs: ```sh curl -fsSL https://claude.ai/install.sh | bash claude --version ``` It needs a Claude account; the first run asks you to log in. If you use Codex instead, everything below works the same way, with `AGENTS.md` where it says `CLAUDE.md`. ## 3. Give it a real task Open a terminal in a folder you actually work in, and start it there: ```sh cd ~/my-project claude ``` Ask for something small and real: "read this folder and tell me what it does", or "this script fails with this error, find out why". Watch what it does before you ask for more. ## 4. The ten things that matter These cover most of what you'll use. Everything else can wait. 1. Start it in the folder you're working on. It reads what's there. 2. Say what you want and why. "Make this faster because the nightly run times out" beats "optimize". 3. Ask for a plan first on anything bigger than a small fix. Shift+Tab cycles the modes; one of them is plan mode. 4. Point at files with `@`, like `@analysis.py`, instead of describing them. 5. Esc stops it. Esc twice lets you go back to an earlier point in the conversation. 6. Read what it asks permission for before you say yes. That is where mistakes get caught. 7. Ask it to check its own work: run the tests, run the script, show you the output. 8. Use `/clear` when you switch to a different task, so the old one doesn't confuse the new one. 9. Commit with git before a big change, so you can always undo. 10. When you're stuck on the tool itself, ask it. It knows how it works, and `/help` lists the commands. ## 5. Your first skill: write like you A skill is a folder with one file, `SKILL.md`, that teaches the agent how to do one thing your way. A voice skill is a good first one, because it's useful on day one and you can tell right away if it works. Create `~/.claude/skills/my-voice/SKILL.md`: ```markdown --- name: my-voice description: Write emails and messages the way I write them. Use for any text I'll send as myself. --- Write like me: - Short sentences. Say the main thing first. - No exclamation marks unless I'm really excited. - No "I hope this finds you well" or other filler. Three things I actually wrote: - (paste a real message of yours) - (another one) - (another one) ``` Start a new session and ask: "use my-voice to answer this email: ...". If it doesn't sound like you, edit the rules and try again. That loop is most of what working with skills is. ## Ready for the next one? Move on to [intermediate](intermediate.md) when you can say yes to all three: - I've used Claude Science or a coding agent on my own research several times, not only tried it. - I've asked for a plan first, and changed the plan before it ran. - My voice skill gets used, and I've edited it at least once after seeing what it wrote. If not yet, keep using it for a week. That's normal, and it's the part that matters.