Chasing a big idea, and letting AI poke holes in it

I have an idea I call Clarus. It is a concept, not a company. I say so out loud whenever I show it to anyone.

The idea is a consultancy that helps local governments modernize: review how each department works, find the waste, propose fixes, estimate the savings, and be paid from savings that are actually proven.

What I built with AI

• A logo, through several rounds, from flat marks to an animated one.
• A website mockup that explains the idea, with a calculator for the cost of doing things the old way.
• A briefing document that works through public budget documents and estimates what a city could save.

Where it got interesting

After writing the first draft of the briefing, Claude reviewed its own work without being asked. It corrected twelve of its own claims.

Some examples of what changed:

• One comparison was dropped because newer data no longer supported it.
• One was rebuilt on a fairer baseline, because the original year had been distorted by a hurricane.
• One proposed saving was removed entirely, because state law requires the expense.

The total estimate went down by about a quarter. A sales pitch that makes its own number smaller is not something I expected, and the document is far more believable for it.

The third draft added a section on what the city already does well. That was the right instinct. Nobody wants to read a report that is only about what they are doing wrong.

The question I have to answer first

Claude also raised something I needed to hear, and raised it more than once. Ethics rules may limit what a public employee can do in business with their own employer. Its advice was to get an opinion from the state ethics board before pitching anything.

That is where the idea stands. It is a concept I am developing, and that question comes before any next step.

What I take from this

Ask the AI to argue against your idea. If it only agrees with you, it is not helping.

Where it falls short

A savings estimate built from public documents is an outside view. It is a place to start asking questions, not a finding.

---

Part of the series 27 ways I actually use AI. Claude drafted this post from the record of our work together.