AI-assisted estimate

Two things go wrong when an AI estimates dev work, and they pull in opposite directions:

1. It anchors on human timelines from training data ("a developer would take 2-3 days") for work an agent writes in minutes. The number comes out several times too high.
2. It forgets the human. Counting only agent minutes misses the part that actually eats the day: reading the diff, running the app, clicking through the flow, waiting on CI, answering review comments, the PM thread. Those are usually most of the real time.

And one thing goes wrong regardless of who estimates: quoting before looking at the code. A number built from a ticket's wording is a guess dressed up as an estimate. It tends to be wrong in whichever direction the code surprises you (a second copy of the form nobody mentioned, a validator only the import uses, a filter built from constants).

The headline number this skill produces is the AI-assisted total: agent time plus human time. That is what the person will actually spend, and it's what gets quoted.

Procedure

1. Lock the scope

Before sizing anything, write down in one or two lines what is in and what is out, as the stakeholder agreed it, not as you'd build it. Read the thread or ticket for decisions already made ("no change to imports", "client wants the basics").

• If you believe something outside the agreed scope is necessary, don't fold it into the number silently. Price it as a separate line and say why. The person decides whether it becomes part of the quote.
• If scope is genuinely unclear and the answer changes the number a lot, give the estimate for each reading rather than guessing one.