How to Edit AI-Generated Prose: The Tells, the Search Terms and the Four Passes

This guide names it. It lists the tells, gives the search terms that find them, and lays out the four-pass edit that turns a model's scaffold into a passage that sounds like a person wrote it, which, by the end, a person will have.

The tells

I keep this list in the banned section of my voice profile and add to it as I notice new habits. Grouped by type.

Structural tells

Phrase tells

Search for each of these. Most should be cut on sight.

a testament to          couldn't help but        in that moment
a sense of              found herself/himself    little did she know
a mix of ... and        the weight of            washed over
a surge of / a wave of / a pang of               palpable
delve                   tapestry                 testament
eyes widened            let out a breath she didn't know she was holding
a reminder that         it was as if             somehow
the silence stretched   time seemed to           her heart raced
a flicker of            unspoken                 the air was thick with
in the end              at the end of the day    navigate (as metaphor)

None of these is forbidden in English. Each of them appears in model output at a frequency that marks the text. If a chapter contains four, the reader's ear has registered it.

Character tells

• Named emotions. "She felt a surge of relief." "Anger flared in his chest." The model tells you the feeling instead of showing the behaviour. See show, don't tell.
• Characters explaining themselves. Dialogue in which people say exactly what they mean and why: "I'm angry because you left without telling me." Real people almost never do this.
• Universal reasonableness. Everyone in the scene is articulate and fair. Nobody is wrong in a way they cannot see.
• Physical reaction inventory. Hearts race, breaths catch, jaws tighten, fists clench. The same six reactions for every emotion in every character.
• The nod and the smile. "She nodded." "He smiled slightly." As beats between every line of dialogue.

Setting tells

• Generic sensory detail: "the smell of old books," "the sound of distant traffic," "golden light filtering through the trees."
• Adjective pairs: "cold, hard stone," "dark, empty streets."
• Weather as mood, every scene.

The search pass

Before any reading, run the mechanical search. In any editor with find, search for each phrase in the list above and for these patterns:

• Emotion nouns: fear, anger, relief, joy, dread, unease, guilt, sadness, hope (as nouns).
• Feeling verbs: felt, realised, knew, wondered, noticed, sensed.
• Adverbs on dialogue tags: find ly," she and ly," he.
• "As if" and "as though."
• Em dashes, if your voice profile bans them; models use them heavily.
• "Something": "something shifted," "something in her chest," "something she couldn't name."

Note the count for each. A chapter of 3,000 words with thirty hits on feeling verbs has a structural problem, not a line problem: the scenes are being narrated from outside. That is a rewrite, not a search-and-replace.

The four passes

Editing AI prose is the same as editing any prose in a particular order, plus the search pass above. Structure first, because there is no point polishing a scene that will be cut.

Pass 1: Structure

Read the generated scene against its scene card. Ask:

• Does the scene start where the card said? Models add arrival: characters walking in, looking around, noticing the room. Cut to the first line where the problem is present.
• Does it end on the turn, or on a reflection after the turn? Models append a closing beat. Delete the last paragraph and see whether the scene is better. It usually is.
• Did the model resolve tension the card said to leave open? Models default to reconciliation: the argument ends with understanding, the threat is deferred. Restore the open question.
• Did it invent a fact? A new object, a new relationship, a piece of backstory. Decide now whether it is true. If not, delete it before it propagates. Our guide to keeping AI consistent explains why "now" matters.

Pass 2: Continuity

Check the scene against the bible: physical facts, who knows what, setting rules, timeline. Models are especially prone to giving characters knowledge from the bible that they should not have in the story: the narrator knows Teo is dying, so Teo mentions it three scenes early. This is a local matching task and the model itself is good at it; paste the scene with the cards and ask for contradictions.

Pass 3: Line

Now the sentences. Working through the passage:

1. Cut the last sentence of each paragraph if it summarises or reflects. Test each; keep the ones that do work.
2. Replace every named emotion with a behaviour, an object the character looks at, or nothing.
3. Break the triplets. Two items or four. Or one.
4. Delete the contrast constructions or convert them to a direct statement.
5. Vary paragraph length deliberately. Merge two; split one into a one-liner.
6. Replace generic sensory detail with a specific one from the bible's setting entry. Not "the smell of old books"; "the smell of the glue they used in the sixties, which had never quite dried."
7. Remove the nods and smiles. Replace with a beat that carries information, or nothing.
8. Cut the phrase tells found in the search pass.

Pass 4: Voice

Read the passage aloud against a sample from your voice profile. The question at this stage is not whether the sentences are correct but whether they are yours. Rhythm, word choice, distance from the character, how figures are drawn. This is the pass where, in practice, most remaining sentences get rewritten, and the pass that separates a manuscript that reads as authored from one that reads as processed.

A worked example

Generated scaffold, from a scene card ("Mara finds the second witness; he refuses to help; ends with her realising he is afraid of Teo, not of her"):

Mara stepped into the small, cluttered workshop, her eyes adjusting to the dim light. The smell of sawdust and oil hung in the air, a testament to decades of careful work. The old man at the bench didn't look up.

"I know why you're here," he said, his voice heavy with something she couldn't quite place. "And the answer is no."

Mara felt a surge of frustration. She had come so far, risked so much. "You're the only one who can authenticate the seal," she said. "Without you, everything falls apart."

He finally met her gaze, and in that moment she saw it: not stubbornness, but fear. It wasn't her he was afraid of. It was Teo.

The realisation washed over her like cold water. She had been so focused on her own mission that she hadn't stopped to consider what her brother had become.

Every tell in the list is present: front-loaded setting, "a testament to," "something she couldn't quite place," named emotions, the contrast construction, "washed over," the reflective final paragraph, and a character (Mara) who explains the stakes aloud to someone who already knows them.

After the four passes:

The workshop had one window and the old man had his back to it. He did not look up when she came in. He was fitting a hinge, and he kept fitting it.

"No," he said.

"I haven't asked."

"You've the look of someone about to." He set the hinge down. "I don't do seals anymore. Not for the archive, not for anyone."

She had planned three arguments and used none of them, because his hands, now they were empty, had started to shake, and he put them flat on the bench to stop it. He was not looking at her. He was looking past her, at the door, the way people look at a door when they are listening for someone on the other side of it.

"He doesn't know I'm here," she said.

The hands did not stop.

The scaffold told me the shape: entrance, refusal, realisation. The edit removed everything that explained, and made the reader do the work of seeing the fear. That is the whole difference between prose that is generated and prose that is written.

What not to do

Do not run it through a "humanizer." Tools that rewrite text to defeat AI detectors produce text that is different but no better, and often worse: they substitute synonyms and scramble syntax without knowing what the scene is for. The fixes above require knowing what the scene is for. Only you know that.

Do not lightly edit. Changing a word per sentence leaves the structure, the rhythm and the tells intact. If a passage is worth keeping, it is worth rewriting.

Do not edit generated prose in bulk at the end. Do it per scene, before the next generation, so that the tells are not fed back into the context and reinforced. A manuscript with sixty thousand words of unedited scaffold is a rewrite, not an edit.

Try it in SparkLore

SparkLore's editing tools include a tell finder that runs the search pass above on a selected passage, with counts by category, and a rewrite action that applies your voice profile's banned list and samples to a scaffold before you do the line pass by hand. The continuity check from pass 2 is one click per chapter. If you want to see the tells in a fresh passage first, the free AI story continuation tool will generate one; run the search list against it and you will find most of the table above in the first three paragraphs, which is the point. Sign up to run the passes inside a project.