Skip to content

Safe start: building with an AI coding assistant

You don't need to know Python well to use midas-nx — tools like Claude Code, ChatGPT, GitHub Copilot, or similar can write the code for you. This page is for that path. It assumes you can describe what you want in plain language but can't fully review the Python that comes back.

The one thing to internalize

Generated code is not verified code, even when it runs without errors. midas-nx drives a live engineering application — a wrong call can discard unsaved work or corrupt a model, and a script that "worked" once isn't proof it's safe to run again. Treat every script an AI writes for you the way you'd treat a script written by a stranger: read what it's about to do before you run it.

The two rules that matter most

  1. The first script should only read data. Don't let an AI assistant's first attempt at your task include create, update, delete, delete_all, new_project, open_project, or analyze calls. Get a read-only version working and understood first, then extend it.
  2. AI assistants guess plausible-sounding function names. midas-nx has changed shape across versions, and an assistant trained on old data (or just pattern-matching) will confidently write calls that don't exist. The context pack tells it not to do this and how to check — use it.

Step 1: give your assistant the context pack

Paste the AI context pack into your chat before asking for any code — as a system prompt if your tool supports one, otherwise as your first message. It tells the assistant the real API shape, the error model, and the specific ways this SDK can hurt you if used carelessly (doc.new_project() discarding work, delete_all() emptying a whole table, timeouts not being rollbacks, and so on).

Step 2: describe the task with this template

Product: Gen NX or Civil NX
Current model state: [describe without confidential information]
Goal: [one concrete task]
Inputs: [ids, load cases, units, selections]
Expected output: [screen output, a summary, or a file]

Start with a read-only version of this. Show me a preview of anything you
plan to create, change, or delete before writing the code that does it.

Step 3: review the generated code before running it

Go through this checklist. If you can't answer a line, ask the assistant to explain it in plain language rather than skipping it.

  • [ ] Does the script only read data (no create/update/delete/ delete_all/new_project/open_project/analyze) — or, if it writes, did I explicitly ask for that after seeing a read-only version work first?
  • [ ] Does it import only things that plausibly exist? (Ask the assistant to show python -c "import midas_nx; print(dir(midas_nx))" output, or point it at the reference, if unsure.)
  • [ ] Is the product (Product.GEN / Product.CIVIL) the one I actually use?
  • [ ] Is my MAPI-Key absent from the code the assistant shows back to me, or from anything it suggests logging/printing?
  • [ ] If it writes data, does it print a preview (ids, payload) before the write call runs?
  • [ ] Is there no delete_all() without confirm=True, and no blind new_project()/open_project() at the top of the script?
  • [ ] If a call can time out, does the script avoid automatically retrying the same write?
  • [ ] Am I running this against a test/disposable project the first time, not my real model?

If something goes wrong: reporting back to the AI

Don't paste raw tracebacks or your MAPI-Key back into the chat. Strip identifying details and use this shape instead:

Installed midas-nx version:
Python version:
Product: Gen NX / Civil NX
Operation type: read / create / update / delete
What I expected:
What happened:
Sanitized error message (no keys, no model/customer data):

Do not suggest an automatic retry. First explain how to safely check
whether the request already took effect.

That last line matters — a timeout is not proof the request didn't happen (see the context pack's safety facts). Letting an assistant "just try again" after a write timeout can double up whatever it was doing.

Next