ChatGPT and Codex training for companies in Frankfurt Rhine-Main

Two tools, two audiences

ChatGPT and Codex come from the same company, OpenAI, and are used by different people. ChatGPT is a chat assistant for writing, research and analysis. Codex is a coding agent: it reads a repository, changes files, runs commands and proposes changes for review.

A training that mixes both loses half the room. I teach them as two tracks. Office teams learn ChatGPT on their own documents. Developers learn Codex on their own code, in a session of their own.

What the ChatGPT track covers

• Instructions that give the same quality twice, and templates the team shares.
• Projects that keep documents and instructions together for a recurring task.
• Checking an answer: sources, numbers and names before a text leaves the company.
• Which data may go into ChatGPT, following the rules of your data protection officer.

What the Codex track covers

• An instruction file in the repository. Codex reads AGENTS.md, and this site keeps its project rules in such a file.
• Limits: which directories the agent may change and which commands it may run.
• Checks the agent must pass before a change counts as done: tests, linters, a build.
• Review: how a developer reads an agent’s change and when a person has to approve.
• Several agents on one code base: Codex and Claude read their instructions from different files, and one file can be the single source that the others import. Details: AI agents under the same rules as humans.

Online or on site

An online session suits teams in several offices and splits into shorter units. An on-site session at your office in Rhine-Main lets developers work on their own machines with access to your repository, and office teams work with your documents in the room. Either way, the exercises come from your company, and the needs assessment decides the format.

A starter instruction file for Codex

What the first AGENTS.md in a repository should answer. Your developers can write it before the training.

1. What the project is, in two sentences.
2. How to install, build and test it, as commands.
3. Which directories the agent must not touch.
4. Which commands need a person’s approval, such as deployments and database changes.
5. The code style, as a link to the linter configuration.
6. What counts as done: which checks pass before a change is proposed.
7. Where the longer documents live, so the file stays short.

Who I am

I am Michael Wutzke, an IT architect for data platforms in Frankfurt with more than twenty years in IT and media. I build software with Codex and Claude in client work, hold a training authorization from IHK Frankfurt and teach at Claude Hacker House. Besides the cloud tools, I am interested in open-source AI models that a company runs on its own servers. Details: Career stages.