The Dumbest Idea I've Had Today (but it's still early)
I was updating this site — trying to replace hypothetical examples with real ones from the build. And then I had one of those ideas that feels stupid until it works (and sometimes still feels stupid after it works).
I've been building with AI for a while now. Tons of projects — some live, some dead repos never to see the light of day. This site, a political stock-trade tracker, an agentic OS framework, a cooking app. Different tools, different AIs, different problems. We're talking weeks of work across each one, context windows stretching back months. The AI didn't remember all of it, but it had enough to see patterns I couldn't.
So I wrote a prompt. Pasted it into four different context windows (thankful for the "undo changes up to this point" feature in Antigravity) and let it cook.
The Prompt
Stop being a coding tool for a minute and be an analyst.
You've been watching me work this entire conversation.
I need to understand my own prompting style — not so I can
tell people to copy it, but so I can teach the PRINCIPLES
behind why it works (or doesn't).
Be brutally honest. Analyze my prompting behavior:
1. What patterns do you see in how I give instructions?
2. What did I do that made your output better?
3. What did I do that made your output worse? Be honest.
4. When I corrected you, how did I do it?
5. Where did I give you autonomy vs. micromanage?
6. What do I do that a beginner wouldn't?
7. Distill my approach into 3-4 principles.
Use actual examples from our conversation.
I want data, not compliments.
What Came Back
They all said the same things. Not word-for-word — but the same patterns, independently identified.
Pattern 1: Lead With Intent, Not Instructions
Every analysis flagged this. I don't tell the AI how to do things. I tell it where we're going and let it figure out the path.
Truthfully, this is because I often don't know the best path forward. I just know what I want the end result to be.
Pattern 2: Correct the Thinking, Not the Output
When the AI gets it wrong, I don't say "that's bad, try again" every time. I reframe the problem.
I'll ask it to think like a user, a developer, or a product manager — whatever perspective it's missing. Often I'll tell it not to make changes yet, just review the problem and find the root cause and work it back. Something like: "Let's pause and step back. High level, this is what I'm trying to achieve. Where are we going wrong?" — then work the response until the AI is back eye to eye with your vision.
Pattern 3: Control Velocity, Not Code
I never tell the AI what code to write. I control what feature or fix we're working on.
AI is always trying to complete as much as it can — it's trying to be helpful. Sometimes it's better to have it do 90% of the work, then go back and narrow in to get the last 10%. Break big tasks into sequential pieces instead of dumping everything into one prompt. Build the database first. Then one integration. Then the next. The AI does better work when it can focus on one thing at a time — same as anyone. (This is exactly what the Scope and Conversation Management guide covers.)
The AI is better at writing code than I am. I'm better at knowing what I want to do with that code — its intent and what the ultimate goal is.
The Others
Three more patterns showed up in most of the analyses:
- Separate diagnosis from action. Don't let the AI fix something it hasn't explained first. "No changes yet, just research." (More on this in the Prompt-Review-Refine Loop.)
- Switch the role, not the prompt. Instead of rephrasing the same question, change who the AI is. "Stop being a coding tool. Be an analyst." "You ARE the LLM we're having issues with — what do you think?"
- Show, don't describe. When words aren't working, paste a screenshot, a log, a reference image. On this site, I pasted a picture of a race car to communicate a color palette. It worked better than any words would have. Snipping Tool and MS Paint are genuinely two of the most useful AI tools I own — one picture can replace a paragraph of instructions.
- Think out loud. If you notice my prompts, I'll sometimes switch from one thing to another seemingly at random. It's not always random. I've found that "thinking out loud" into the prompt gets better results — it gives the AI more context about what I'm actually thinking, which makes its suggestions better.
The Part Nobody Said Out Loud
Every analysis — all four of them — ended at the same place. Different words, same conclusion:
The skill isn't in the prompt. It's in the feedback loop.
My prompts aren't polished. They're full of typos and mid-thought pivots. What makes them work isn't the opening message — it's what happens between messages. The corrections. The redirects. The instinct for when to let the AI run and when to pull it back.
Most people think the goal is writing the perfect prompt. The actual goal is learning to correct fast and correct well. Three rounds of specific feedback beats one "perfect" prompt every time.
Why I'm Telling You This
These patterns aren't rules. They're just what I figured out by doing, confirmed by asking the tools themselves. If you're trying to get better results from AI, they might save you some of the trial and error.
That's the whole point of this site. I hope it helps.