The single most frustrating thing about working with AI isn't when it says "I don't know." It's when it hands you something polished, confident, and quietly wrong — a statistic that doesn't exist, a summary that misses the one clause that mattered, a plan with a step missing in the middle.
There's a simple habit that catches a surprising amount of this, and it costs you one extra sentence: make the model check its own work before it answers.
Why confidently wrong happens
A language model is optimized to produce a plausible next piece of text, not to verify it. Fluency and accuracy are different things, and the model doesn't feel the difference between them. So when it doesn't know something, you rarely get hedging — you get the same confident register as when it does know.
That means the burden of review lands on you. The self-check pattern moves part of that burden back where it belongs.
The scaffold
Instead of asking for output and accepting the first thing you get, build critique into the same request:
- Ask for the output.
- In the same prompt, tell it to critique its own draft against criteria you define.
- Have it revise, and only then return the final version.
The bare-bones version you can copy right now:
Draft [X]. Then list 3 ways this could be wrong or weak. Revise to fix them. Return only the revised version.
That's it. The reason it works is that criticizing a finished draft is an easier task than producing a perfect one, so you're asking the model to do the thing it's comparatively better at.
Make the criteria specific
"List 3 ways this could be wrong" is a decent default, but generic critique produces generic fixes. The pattern gets much stronger when the criteria come from you, because you know what actually goes wrong in your work.
Compare:
- Generic: "Check this for problems."
- Specific: "Check this against three things: (1) every number is traceable to the source I pasted, (2) no claim promises a specific result, (3) the tone matches the sample email above. List each failure, then revise."
The second version catches the mistakes that would have cost you something. The first mostly produces reassurance.
A practical way to build your criteria: the next three times you edit AI output by hand, write down why you edited it. Those reasons are your checklist.
Where the self-check genuinely helps
- Anything with numbers or citations — forcing traceability catches invented figures.
- Summaries of long documents — "what did you leave out that changes the meaning?" is a great second pass.
- Multi-step plans — asking it to find the missing step catches gaps before you execute them.
- Anything customer-facing — tone and over-promising checks are cheap insurance.
Where it doesn't help — and this matters
Be honest about the ceiling here. A self-check reduces error; it does not eliminate it. If the model confidently invented a fact, it can just as confidently confirm that fact during its own review — the same blind spot produces both. Self-critique is weakest exactly where the model's knowledge is weakest.
So the rule is: use the self-check to raise the floor, not to skip verification on anything that actually matters. If a number goes in front of a customer, a regulator, or your bank, you check it yourself. The pattern buys you fewer obvious misses, not a guarantee.
From a habit to a workflow
Once you've run a self-check manually a few times and you trust the criteria, that's the moment it's worth automating. A loop — plan, act, self-check, stop — is just this habit with the review step built in permanently, so you don't have to remember to ask.
That's the architecture behind the Prompt Loops Vault: eight autonomous work loops that plan, act, check their own output, and stop when the job is done. If you'd rather start smaller and free, the 10-prompt sampler is a good place to practice the pattern, and the difference between a prompt and an agent explains when it's worth graduating.
These are tools you customize to your own context, not guaranteed outcomes. But of all the habits worth building, making the AI argue with itself before it argues with you is the cheapest one that actually pays.