Decide Whether AI Output Is Good Enough to Use
AI output usually reads smoothly. The problem is not fluency; it is which sentences survive a check. This workflow sorts first, then decides what to cut, fix, or keep.
How to do it
Step 1: Classify first, do not edit yet
Use “Separate Fact, Inference, and Preference” to classify each sentence: a fact a third party could verify, an inference drawn from facts that could be wrong, or a preference that is a stance. Then find the sentence used as fact that is really inference. Only verifiable sentences need a source; opinion sentences need to earn their place.
Step 2: Run “Fact and consistency check” line by line
Ask for three labels: contradicts another part of the text, contradicts common knowledge, or cannot be confirmed. Fix the first two on the spot. Do not delete the third; mark it as unverified and note where it would be checked. Unverified is not the same as wrong, but it is not ready to publish either.
Step 3: Test the evidence with “Check whether claims match their evidence”
Use “Check whether claims match their evidence”: data with too small a sample, a single case presented as a pattern, and a citation with no source all count as a mismatch. When the evidence is thin, soften the wording or go find the source. Never use a more confident tone to cover weaker evidence. One “possibly” costs nothing; one invented source discredits the whole piece.
Step 4: One last attack with “Red-Team My Judgment”
Use “Red-Team My Judgment” for one last attack: find the three conclusions that hold up worst, and ask for a counterexample or an alternative explanation rather than a vague “needs attention”. The goal is not to rewrite the piece beyond recognition but to confirm that every remaining conclusion survives one honest objection. If it does not, downgrade it to “possibly” or cut it.
Step 5: Finish with “Strip the AI voice”
Use “Strip the AI voice” for the final pass: cut parallel constructions, signpost words like “first, second, finally”, and replace abstract nouns with concrete actions. Read it aloud; the place that sounds awkward is usually the place you have not thought through. If a sentence collapses when you remove the filler, it had no information to begin with.
Prompts used
FAQ
Can I use numbers an AI gives me?
No. Numbers, regulations, citations, prices, and version numbers all go back to the original source. If you cannot find it, cut the number or mark it clearly as unverified.
How do I spot the part the AI made up?
Look at specificity: precise numbers with no source, names with no link, and causal chains that run too smoothly. Classifying each sentence with “Separate Fact, Inference, and Preference” is the fastest filter.
If I rewrote AI output, is it still AI-generated?
That depends on the use and the platform's rules. What is certain is that you are responsible for every sentence you keep, so checking and rewriting are not optional.