FAQs

AI-Assisted Programming for PhD Researchers

Answers to the questions we hear most often. If yours isn’t here, ask the Oshu chat assistant in the corner of this site, or email me. For install questions, start with the Setup page.

How do I see the slides?

Each lecture links to its slides (a RevealJS deck) from the top-right of the lecture page. In the deck, press E to switch to a print-friendly layout, then use your browser’s Print → Save as PDF to export a copy for offline reading or annotation.

Do I need to pay for anything?

No. Mistral’s free “Experiment” API key from Setup (one key powering both the agent and edit predictions) covers everything the course requires. The Zed student plan is a nice optional extra, and you may also bring a subscription you already pay for (a paid Mistral tier, ChatGPT, a Claude subscription, or an Anthropic API key). You only incur costs if you choose to use a paid option.

What if my Zed student verification hasn’t come through (or I skipped the plan)?

You are not blocked: the student plan is a recommendation, not a requirement. Edit predictions and everything else in the course run through your free Mistral API key (Setup, step 2). The plan only adds Zed-hosted extras ($10/month model credits, Zed’s own predictions).

What if I hit Mistral rate limits during class?

Tell me. A backup key pool exists for use during the workshop, so a rate limit on your personal key won’t stop you from finishing a lab.

Can I use Zed’s built-in agent instead of OpenCode?

Yes. Zed ships its own agent in the agent panel, and it has become genuinely good. It runs on a Mistral API key you paste into Zed’s agent settings (the same Experiment key from setup, in its own key slot), or on student-plan credits, and the labs’ prompts and workflow (plan, act, review the diff) work the same. You can also run OpenCode inside Zed: the agent panel speaks ACP, and OpenCode installs from the ACP registry, which gives you OpenCode’s sessions with Zed’s inline diff review. The lab instructions and cheatsheet name OpenCode’s UI (Tab for plan/build, /undo), so if you use Zed’s own agent you translate those details yourself.

Can I use Claude Code, Copilot, or Cursor instead?

For your capstone, yes: use whatever agent you prefer. The labs are written for OpenCode, so following along is easiest with it, but the concepts (plan/build modes, context files, the writer/reviewer pattern, MCP) transfer directly to other agents. One caveat: capstone Options A and B rely on the SKILL.md skills built in Lab 5, and OpenCode and Claude Code can both load these directly. Copilot and Cursor cannot use them the same way, so if you plan to use one of those tools, pick a capstone project that doesn’t depend on the skills, or plan to translate them yourself.

What data may I send to the model?

  • Course lab data: yes. The lab dataset is synthetic and safe to share with the model.
  • Your own unpublished research data: only if you have checked your own data-protection constraints first (funding terms, ethics approvals, personal or sensitive data). We cover this in Lecture 06. When in doubt, use a synthetic sample instead of real data.

Is the text and code I generate watermarked?

Increasingly, yes. Under the new EU AI Act (Article 50, applicable since August 2026), providers of generative AI have to embed machine-readable marks in their content. Claude models released since August 2026 watermark all text output (including code) and Google’s Gemini has done this since 2024 (SynthID). A watermark proves that a model processed the text, not who did the work, and heavy editing can remove it, therefore its absence proves nothing either. The practical consequence: disclose your AI use rather than rely on it going unnoticed. We cover the mechanism and the disclosure norms in Lecture 06.

How does the chatbot work?

The Oshu assistant embedded on this site answers questions from the course content. It runs on custom code hosted on Hetzner in Germany and generates its answers through the EU API of Mistral. Your messages are not used to train AI models. See the Privacy Policy for the full details.

Do I need to know Python?

Basic scripting experience in any language is enough. The labs use Python, but every step is guided, so you will not be left to write Python from a blank page. The course is about working with an agent, not about Python syntax.

What do I have to deliver?

Nothing that is graded. The course is pass/fail on attendance, so being there for the three days is what you have to do.

You will build a project of your own on Day 3 and show it to the group in the afternoon: run it, say how you used AI on it, and tell us what went wrong. That is for the room’s benefit, not for a mark. Details are on the Showing Your Work page.