#090 - Managing AI Tools: A Framework for Automated Design | 01 Context and Compaction

How To Maximize Efficacy of LLM's for Professional Engineering

Large language models fail on complex engineering problems because adequately contextualizing multi-variable design work is genuinely difficult. The model has no memory of last week’s meeting. It doesn’t know which version of a structural code your company or your client has adopted, which constraints actually bind the problem, or what failures cost you before. You can dump all of that into a prompt, but there’s a mechanical limit to how much context a model can process before reasoning degrades. Past that threshold, you’re adding noise.

For a greenfield task, writing a standalone Python script or drafting a generic email, LLM’s are incredible time savers. Point that same tool at a complex, real-world project with ten years of history, conflicting geotechnical reports, and specific client requirements or code amendments, and things begin to unravel. The model hallucinates code provisions (less so with the more recent models thankfully). It ignores constraints you explicitly provided minutes ago. You end up with a massive volume of output that looks like progress but is actually rework and technical debt in disguise.

This is a massive risk and ongoing problem for less experienced engineers who are trigger happy with AI tools. You don’t know what you don’t know. A little knowledge is a dangerous thing, and a language model will happily amplify this at scale.

In software development, this reliance on loose prompting and hope is called “vibe coding.” In our industry, it’s essentially the same thing. It’s inefficient, it compounds errors, and it relies on the false hope that the AI will simply figure it out without structure.

The lesson is simple: do not outsource the thinking.

If you are familiar with prompt engineering and context engineering, please read on, if not. I absolutely recommend you check out this previous article for an important primer to get the most of the sections below…

#053 - Effective Use of Large Language Models in Engineering