If you've been using AI more in your role, you've probably tried feeding Claude or ChatGPT a few past projects and asking it to design the next one with similar parts. It works well enough to be tempting, and poorly enough to be risky.
The model has to work out your procedure from examples, so it works it out differently every time. The more projects you paste in, the more it loses track of the ones in the middle. And it picks parts from its training data, not current documentation.
The goal is an AI design system that produces consistent, accurate, repeatable results and gets better with every project your team completes. The way to get there is to stop giving the AI examples and start giving it instructions.
That takes four steps, done in order.
A skill is a written procedure for one task. Start with the task you do most, such as a standard conference room. Write it the way you'd train a new engineer:
The inputs it needs: room size, seating count, display count, client standard. The decision rules: display size by viewing distance, DSP by mic count, and so on. Your approved parts list. The output format you want. Then test it.
Run the skill on a past project and compare the result to what you actually built. Wherever it differs, fix the skill. Then move to the next task.
This works because the agent follows your procedure instead of guessing, and it only loads a skill when the task calls for it, so your context stays clean.
Keep them in one location that every engineer's agent reads from, with one person approving changes. When someone finds a mistake, they fix the skill, not their own prompt. If each engineer keeps private skills, you've just moved the inconsistency from the model to your people.
Set up a dedicated machine in your office to run the agent. The model itself still runs in the cloud; the machine holds the agent, its skills, and its files. Work continues when your laptop is closed, and everyone connects to the same agent with the same skills.
Put your projects, products, and client standards in one database on that machine, and connect the agent to it so it can read and write. Also give the agent access to manufacturer documentation, so it checks compatibility against current specs instead of its memory. At the end of every project, have the agent write back the final parts list and lessons learned.
Every job then makes the system smarter without making your prompts longer.
The order matters.
Skills are the foundation. A dedicated machine without skills is just an expensive computer, and a database without skills is data the agent doesn't know how to use.
Build the skills first.