Newsletter
Your AI writes beautifully. Your reports still bounce before they go out
Your AI writes well, yet your reports still bounce. The gap is that nobody wrote down how your firm writes. How we fixed that, for engineering firm owners.
For engineering firm owners
The gap was never the writing. Nobody ever wrote down how your firm writes.
Updated November 2026 from my July letter. What changed: how our skills work today, why we took the AI out of my own example, and what we have not measured.

A senior engineer on my team was reviewing another firm's draft of a permit application. Within a few minutes he called it. It read like a third-year EIT had assembled it. Half the forms were missing. It met the minimum and stopped there. He was clear it was a good firm. But you could see who built it before you ever saw the name on the cover.
That is just as true inside your own walls. Your best PM turns in a basis-of-design report that reads tight. The new EIT turns in one that is technically correct and structurally a mess, and you already know it will bounce off your reviewing PE before it reaches the client. The engineering underneath both is fine. The difference is how it was written.
When I first put AI on our narrative deliverables, I assumed the writing would be the problem. Early on it was wordy. That turned out to be the easy part. It read like a competent stranger had written a report about water infrastructure, because that is exactly what happened. The model could write. It had no idea how we write.
AI does not have a writing problem. It has an instructions problem.
How your firm writes lives in two or three heads. Nothing a machine can read.
Your firm has a way of structuring a report and a tone for a response letter. None of it is written down. It lives with the people who learned it over fifteen years, and it leaves when they do. Until then, it is why your senior engineers rewrite drafts instead of reviewing them.
Polished prose can make that worse. BetterUp Labs surveyed 1,150 full-time US desk workers in September 2025. Four in ten said they had received AI-made work in the past month that looked finished and wasn't, and each case took about two hours to sort out. It is self-reported, from a coaching company, but it matches what I saw: a fluent draft with the wrong structure costs a reviewer more, not less.
So we write it down, one skill per kind of deliverable, read by our AI agents. In July I described an eleven-part template. We dropped it. A skill now opens with two lines: a trigger in plain words saying when it applies, and who may use it. Then the instructions: what good looks like, what to read first, and our judgment calls. 63 of our 78 published skills follow that shape. When a request matches a trigger, the agent loads the skill, and it is told to say which one it loaded.
Someone in the department that owns a skill can change it in a conversation. They see the exact change and say yes. The next request uses it, with no deploy and nothing to clear. Every change keeps a version, and two people can't silently overwrite each other.
The example I used in July, we took the AI out of.
Some sentences are a statement your PE stands behind.
My July example was an engineer asking the agent to help draft a response-to-comments letter. We changed that rule. The engineer now writes every response. The tool builds the letter around those words. It quotes each reviewer comment word for word, in the reviewer's numbering. It prints the engineer's answer under it, flags any comment with no answer, and marks words like "complies" or "meets the requirements" for the engineer to confirm. It has no AI in it.

Why? A response letter is a record. When a reviewer asks a year later what we told them about a detail, the answer has to be the engineer's own words. The skill for that letter spells out what the machine may never write. That is still writing down how we write.
Now the honest part. In July I said your newest engineer's first draft would start where your best engineer's used to. I have not measured that. Our standard names first-round pass as the test, and we have no before and after. I also said anyone close to the work would add to the library. So far, of 423 recorded changes to our skills, 85 came from people, and those came from five of us. The rest were setup scripts, publishing from our code and automatic updates.

What I can show you is where to start. We counted 10,884 comments our reviewers wrote on site plans across 131 sites. The same twenty items came back again and again, each on at least 51 sites. That list is now the designer's checklist before review. Your reviewers' markups are the best record you have of what your firm never wrote down.
The AI is not writing the report. Your engineers still own the judgment, the numbers and the seal. Every correction someone bothers to write down makes the rest better.
Here is the exercise. Pick the deliverable your senior people keep quietly rewriting. Pull the reviewer markups on the last ten of them and list every comment that shows up three times or more. That list is your first skill.
Tell me which deliverable it is, and the comment that keeps coming back through our contact page. I read every one.
Michael Groselle, PE
Founder and CEO
Modern Engineering Solutions
mod-eng.com
Keep reading
More to read
- Your project manager inherits the files. Everything that actually matters stays with whoever sold the work
- 192 hours worked. $50,000 short on billing. The problem was not effort
- I spent eight rounds making a blind tool smarter. A plain chat beat it, so I closed it
- I told you to audit the setting. I didn't audit mine
What we do
Planning a project?
Talk to an engineerTalk to an engineer
Tell us about your project
An engineer reads it
Not a salesperson. Someone who can tell you whether the pathway is real.
You get a first answer
Your options, the agencies involved and the questions that decide cost and schedule.
You decide
No commitment to have the first call.
Prefer the phone? (214) 833-6748