Ipromptism 檢查清單：保持 AI 素養的 7 個習慣

A few months ago I published Ipromptism: Is AI Creating a New Form of Illiteracy?. The idea was simple: we are learning to prompt faster than we are learning to think with AI. We can get a fluent answer in seconds, yet many of us are losing the ability to judge, verify, or rebuild that answer ourselves. I call this pattern Ipromptism (the framework is archived on Zenodo: 10.5281/zenodo.23212581).

The most common reaction I got was: "OK, I recognise the problem. Now what do I actually do on Monday morning?"

This post is my answer.

Why a checklist?

I have trained more than 2,500 people on AI tools, from developers to managers to people who had never opened ChatGPT. One thing comes up in almost every session: awareness alone does not change habits. People nod at the concept and then go back to copy-pasting the first output they get.

Checklists work because they are boring in the best way. Pilots and surgeons use them because good intentions break down under pressure. Shipping code with an AI assistant at 6 p.m. on a Friday is exactly that kind of pressure.

So here are seven habits I recommend to builders who want to use AI heavily without outsourcing their thinking.

The Ipromptism checklist

1. Write the goal before you write the prompt

Before you open the chat window, write one or two sentences in your own words: what problem am I solving, and what does "done" look like?

• If you cannot state the goal without the AI, you are not ready to prompt yet.
• Keep this note next to your prompt. It becomes your yardstick for judging the output.

This one habit turns the AI from an oracle into a tool you are steering.

2. Predict the answer first

Take thirty seconds to guess what a good answer should contain: the approach, the main risk, the rough shape of the code.

• If the output matches your prediction, you have gained confidence.
• If it does not, you have learned something, or caught a mistake.

Either way, you stay an active participant instead of a passive reader.

3. Verify, don't vibe

Fluent text feels correct. That feeling is not evidence.

• Run the code. Write or run the tests.
• Check every API, flag, and library version against the official docs. Hallucinated parameters look perfectly plausible.
• For factual claims, open the source. If there is no source, treat the claim as a hypothesis.

A good rule: the more confident and polished the answer sounds, the more carefully you should check it.

4. Never merge what you can't explain

This is the anti copy-paste rule. Before AI-generated code goes into a commit, ask yourself:

• Could I explain each line to a colleague in a code review?
• Do I know why this approach was chosen over the obvious alternative?
• Would I be able to debug it at 2 a.m. without the assistant?

If the answer is no, slow down. Ask the AI to explain itself, simplify the solution, or rewrite the tricky part yourself. Code you don't understand is technical debt with a friendly face.

5. Keep the human decisions human

AI is great at generating options. Choosing between them is still your job, especially when the choice involves:

• Security, privacy, or user data
• Architecture and long-term trade-offs
• Anything that affects real people (hiring, access, pricing, moderation)

Use the AI to widen the set of options and to stress-test your reasoning. Then make the call yourself, and be able to say why.

6. Document your prompts like you document code

Good prompts are reusable assets, and bad ones are silent sources of bugs.

• Save the prompts that work in your repo or team wiki, with a short note on context and limits.
• Version them when you change models or tools; the same prompt can behave differently.
• When an AI-assisted change ships, mention it in the PR description. It helps reviewers know where to look harder.

Documentation also forces you to understand why a prompt works, which is literacy in action.

7. Teach it forward

The fastest way to fight Ipromptism in a team is to talk about it openly.

• Pair on AI-assisted tasks the way you would pair on code.
• Share failures, not just wins: the hallucinated function, the subtly wrong SQL, the confident but outdated answer.
• Help juniors build fundamentals alongside AI, not instead of it. They need to know what good looks like before they can judge what the model produces.

A team that discusses how it uses AI will catch more errors than a team where everyone prompts alone.

A quick self-check

If you want a one-minute version, ask yourself after your next AI-assisted task:

• Did I write the goal before prompting?
• Did I verify the output against something real?
• Could I explain and defend the result without the AI?

Three yes answers means you are using AI as an amplifier. If you answered no to any of them, that is exactly where to start.

Closing thoughts

AI is not making us illiterate by itself. The way we use it can, if we keep trading understanding for speed. The good news is that the fix is not complicated. It is a handful of small, repeatable habits that keep your judgment in the loop.

I work on these questions every day as an AI trainer and consultant, with teams and organisations that want to adopt AI without losing their edge. If you'd like to go further, you'll find my workshops, talks, and resources on romainbailleul.fr.

I'd love to hear from you: which habit on this list is hardest for you to keep, and what would you add? Let me know in the comments.