ChatGPT and Codex training for industrial companies in Munich

Where the work piles up at a supplier

At an industrial supplier, work piles up in two places: in the inquiries that sales has to answer and in the old code that IT has to keep running. ChatGPT can take on part of the first, Codex part of the second. Both need a person who checks.

An inquiry arrives with a customer specification that can run to dozens of pages. Someone in sales or in the engineering office reads it, lists what the customer requires, asks back where it is unclear and writes the offer. ChatGPT can do the first reading and the first draft.

In IT the pile is code, such as an add-on to the ERP system or a tool that prepares test data. Such software may be many years old, and its author may have left the company. Codex can read that code and change it. Without rules it changes too much at once.

The ChatGPT track: specifications and offers

Sales staff and the engineering office learn ChatGPT on inquiries your company has received.

• Requirements from a specification as a list, each with its page reference, so a person can check it.
• Two revisions of a specification compared, with the changed requirements listed.
• Questions back to the customer where the specification leaves a gap.
• The offer text from your calculation and your standard terms. The prices come from your calculation.
• A project per product line that holds your data sheets and the terms your company uses.

The Codex track: changes to legacy code

Developers learn to let Codex read before it writes. It explains a module and lists its callers, writes tests that pin today’s behavior and then makes one small change per pull request. A change to the database schema or to an interface with another system waits for a person’s approval. The checklist below puts these rules into the AGENTS.md of your code base. How a machine enforces such rules is described on AI agents under the same rules as humans.

An AGENTS.md checklist for legacy code

What the instruction file of an old code base should state. Your developers can fill it in before the training; each blank shows where knowledge is missing.

1. The language and framework versions the code needs, and how to install them.
2. The build and test commands that work today, even when they are slow.
3. The modules that stay frozen, and the reason.
4. The interfaces to other systems, such as the ERP or machine controls, and who approves a change to them.
5. Tests that pin current behavior come before any change.
6. No dependency upgrade in the same change as a fix.
7. No reformatting of files the change does not need.
8. The person who knows the code best, and where the old documentation lives.

Who I am

I am Michael Wutzke, an IT architect in Frankfurt who builds software with Codex and Claude for a client. In more than twenty years in IT and media I was project manager for the web portals of an insurance group and led projects for online shops at an e-commerce company. In both jobs I defined the scope of projects and challenged other people’s concepts. I teach at Claude Hacker House and hold a training authorization from IHK Frankfurt. I am also interested in open-source AI models that a company runs on its own servers. Details: Teaching and certifications.