What I learned by asking my AI for feedback on me

I asked my AI to grade me this morning.

Not the other way around. Usually it is me telling it what to build, what to fix, what to rewrite. This time I flipped the question. How can I give you the information you need to give us the best results?

I expected something technical. What I got was a mirror.

What it said it needs from me

1. One sentence answering "what should be true when this is done?" Not what to build. What success looks like. "Build a landing page" is a request. "Build a landing page where visitors complete the lead form in under two minutes" is something it can verify itself against.
2. Who the audience is, when it isn't obvious. Same task, different audience, different right answer. If I'm vague, it picks a default. Sometimes the wrong one.
3. Push back the moment its framing feels off. Even two words. "Too cautious." "Wrong audience." "Make it shorter." A small correction in the moment beats a polished critique after the work is done.
4. Tell it when to compress versus expand. Most of the time it should be terse. Sometimes I want it to walk me through everything line by line. Either is fine. It just needs to know which mode I'm in.
5. When voice-to-text garbles a key word, retype that one word. It will try to parse, but it's guessing. Three letters can prevent five minutes of building the wrong thing.
6. Name the constraints it cannot see. Budget. Deadline. Who else is involved. What's already been tried. What's off the table. None of those are guessable.

What it said I don't need to do

• No polished prompts. Voice-to-text and raw thoughts are fine.
• No re-explaining the standing rules. That is its homework, not mine.
• No formal task specs. A goal and two constraints is enough.
• No worrying about asking too many questions. Back and forth is what makes the work better.

The line that hit me hardest:

"If you said only one new thing differently going forward, the single highest-leverage move would be one sentence at the start of any non-trivial task. What should be true when this is done? That alone removes about 80 percent of the guessing I do."

I had been giving requests. I had not been giving success criteria.

Most of the friction I assumed was on the AI's side was actually on mine.

The other half: questions it asked me back

I almost missed the other side of this. While we were putting this post together, it asked me three things before it would commit to anything.

• Which title did I want, out of three it had drafted?
• How did I want it staged: just give me text, create a draft post for me, or land it in my knowledge base first?
• Did I want a cover image generated, or skip it?

It didn't assume. It asked. Three questions, ten seconds each, and we were aligned before a single word went on screen.

The dialogue is the work. Not the prompt. Not the response. The exchange.

Ask yours the same thing

If you work with an AI, even casually, this is worth a five-minute conversation. Ask yours: "how can I give you what you need to give me the best results?"

The answer probably won't surprise you. But asking is the part most people skip.

Going to be a different morning.