How Listening to AI Can Help.
When I first started working with AI, my workflow was:
- Write a prompt.
- Hit generate.
- Hope for the best.
- Patch obvious problems with more words and more constraints.
If the output was wrong, I assumed the model needed more detail. So I added detail. When that failed, I added even more. While my prompts grew longer, the results got worse.
Here’s what I do now
Many LLM-based tools now expose some form of “reasoning” or internal commentary as they work. They don’t just give you an answer; they show you how they interpret your request in real time.
That internal monologue is a goldmine.
Take Claude Code as an example: If you turn on “plan more” or “ask before edits,” you get full transparency into how it understands your instructions before it executes the plan. Which means you can spot misinterpretations early, correct them in plain language, and only then, let it run.
That alone saves tokens, time, and a lot of frustration.
If you try this approach and you’re still not getting the right results (or if you let it auto-edit, which I do not recommend), use the output as a lesson:
- Scroll up and read how the AI described your request.
- Copy its own explanation of what it thinks you asked for.
- Use that language to rewrite your next prompt.
- Start a fresh chat (/clear) and combine what you actually want and what you now know the model will infer.
Over time, your prompts become sharper, and so does the output.