The Problem Isn't Telling. It's Telling Twice.
Every writer has heard "show don't tell" so many times it's become background noise. But that advice assumes the draft is only committing one sin at a time — either it shows, or it tells, and you just need to pick the right one. AI drafts rarely work that way. They tend to do both, back to back, like they don't trust the first attempt to land.
Here's what I mean. The AI will write a tense little exchange where a character's hands shake and their voice cracks mid-sentence. Good stuff, genuinely dramatized. Then, immediately after, it adds: "She was terrified." You already knew that. You watched it happen four lines ago. The narration isn't adding information — it's double-checking that you received the memo, like a text message followed by a phone call confirming you read the text message.
This is a distinct failure mode from generic telling, and it needs its own diagnostic pass. If you only hunt for "telling" in isolation, you'll miss these pairs entirely, because each half looks fine on its own. The dramatized beat is often genuinely well-written. The summary sentence, read alone, isn't even bad prose. The problem only exists in the relationship between the two. You can't catch it by scanning for weak verbs or adverbs — you have to catch it by noticing redundancy across sentence boundaries, which is a harder, slower kind of reading.
The dramatization did the work. The narration is just there to make sure you noticed the work being done — and that's the insult.
What the Pattern Actually Looks Like on the Page
Once you know what you're looking for, you'll see it everywhere in an unrevised AI draft. The shape is almost always the same: action or dialogue that dramatizes an internal state, followed by a flat declarative sentence that names the exact emotion or fact just performed.
- Emotion version: "His fork clattered against the plate. 'I'm fine,' he said, not looking up. He was clearly upset."
- Fact version: "The ledger showed three missing shipments in a row, all signed off by the same clerk. Someone inside the company was stealing."
- Relationship version: "She reached for his hand under the table before either of them had said a word. They'd always understood each other without needing to talk."
Notice the structure is identical across all three. Dramatized beat, then a sentence that functions like a caption underneath a photograph nobody needed captioned. The tell sentence is usually short, often starts with "She was" or "He felt" or "It was clear that," and almost always could be deleted with zero loss of information — because the information was never the point. The scene already delivered it.
This habit comes from how language models are trained to be helpful and unambiguous. They don't want to risk you missing the emotional beat, so they over-deliver: dramatize it, then spell it out. It reads like a safety net the story doesn't need. If you're working through a full manuscript, this is exactly the kind of pattern worth hunting during a dedicated pass — see The Five-Pass Revision Order for AI-Assisted Novels for where this fits alongside continuity and pacing passes.
Why Generic "Find the Telling" Prompts Don't Catch This
If you ask an AI editor to "find places where I'm telling instead of showing," you'll get a list of adverbs, filter words, and abstract emotion-naming sentences flagged in isolation. Useful, but it misses the doubling entirely, because the model is evaluating each sentence on its own merits rather than checking it against the sentence before it. A tell sentence that follows a strong dramatized beat and a tell sentence that stands alone with no setup are two completely different problems requiring different fixes. Lumping them together wastes your revision time on the wrong cases.
You need to prompt specifically for the pairing, not the telling. That means asking the AI to look backward from every "naming" sentence and check whether the preceding paragraph already dramatized the same content. This is a comparison task, not a classification task, and most default editing prompts don't ask for that comparison at all.
Read the attached chapter and find every instance where I dramatize an emotion or fact through action, dialogue, or sensory detail, and then immediately follow it with a narration sentence that explicitly names that same emotion or fact. For each instance, quote the dramatized beat, quote the redundant tell sentence that follows, and explain in one line what information the tell sentence adds that the dramatization didn't already deliver. If it adds nothing, say so plainly. Ignore sentences that tell without a preceding dramatization — I only want the paired redundancy, not telling in general.
This works because it forces the model to hold two sentences in relation to each other rather than judging either in a vacuum. The instruction to explicitly ignore standalone telling is doing real work here — without it, you'll get your list polluted with unrelated issues that belong in a different pass. If you want to narrow this further, you can scope it to a specific concern:
Scan this scene specifically for moments where a character's fear, jealousy, or guilt is shown through physical action or subtext in dialogue, and then restated by name in the narration within the next one to two sentences. List each pair. Then rank them by how much the tell sentence undercuts the tension of the scene — a tell that kills a suspenseful moment should rank higher than one in a low-stakes exchange.
Ranking by damage is the useful part. Not every redundant pair is equally costly — one buried in a quiet transitional paragraph barely matters, but one sitting right after your best line of dialogue in the climax is actively sabotaging your strongest material. Prioritizing lets you fix the worst offenders first if you're short on revision time. For a broader pass that catches filter words feeding into this same habit, the Filter Word Finder is worth running alongside this prompt, since filter words like "she felt" or "he realized" are often the hinge word connecting the show to the redundant tell.
Three Ways to Cut It — and When to Use Each
Once you have your list of paired redundancies, you've got three real options. They're not interchangeable — picking the wrong one for a given scene can flatten a moment that needed the extra beat, just not in its current form.
1. Delete the Tell
This is the right call maybe 80% of the time. If the dramatization genuinely lands — the shaking fork, the cracked voice, the hand reaching under the table — the explanatory sentence is pure scaffolding and comes out clean. Read the paragraph without it. If nothing is lost, it wasn't load-bearing.
Here's a paragraph from my manuscript. Rewrite it by deleting the final sentence that names the emotion, and instead extend the dramatized action by one beat to make the emotional state land even more clearly without naming it directly. Keep the same point of view and voice. Here's the paragraph: [paste paragraph]
Asking for an extra beat instead of a flat deletion matters — sometimes the tell sentence was compensating for a dramatization that was slightly underpowered, not perfectly fine. This prompt fixes both problems in one move instead of leaving you with a technically-correct-but-thin paragraph.
2. Delete the Show, Keep the Tell
Counterintuitive, but sometimes right. If the dramatized beat is clunky, over-long, or slows a scene that needs to move fast, and the "tell" sentence is actually doing the job more efficiently, cut the dramatization and keep the summary. This is especially true for minor emotional beats in high-velocity scenes — a thriller chase sequence doesn't need three sentences of physical tension-building if "he panicked" gets the reader moving just as fast. If you're writing in a genre where pace is the whole game, this is worth remembering — the thriller with AI workflows benefit a lot from aggressively trimming the dramatized half in favor of momentum.
3. Split Them Across Time
This is the most sophisticated fix and the one AI is worst at suggesting on its own — you usually have to ask for it directly. Instead of deleting one half, you separate the show and the tell by distance: the dramatization stays where it is, and the naming of the emotion gets moved several paragraphs or even a full scene later, often reframed as the character's own delayed realization rather than the narrator's immediate gloss.
In this scene, the character's grief is dramatized through her silence at the funeral, but then immediately named in the next sentence ("She was devastated"). Instead of deleting that sentence, move the explicit naming of her grief to a moment later in the chapter — ideally when she's alone and some small ordinary object triggers a delayed emotional reaction. Write that later moment as a short paragraph, and remove the premature naming from the funeral scene entirely.
This technique turns a redundancy into a structural asset — delayed emotional recognition is a real craft tool, and it reads as intentional pacing rather than a cut you made because an editing pass told you to. It's also a good move to pull from if you're working on a Romance Beat Sheet: Where AI Drafts Usually Break scenario, since romance leans heavily on delayed emotional payoff for its best moments.
A Worked Example: Before and After
Here's an unedited AI-drafted paragraph, the kind you'll recognize instantly if you've been writing a book with AI for more than a chapter or two:
Marcus set the phone down carefully, like it might go off again if he moved too fast. His jaw was tight, and when he finally spoke, his voice came out flatter than he intended. "They're not coming." Elena watched him for a long moment before she said anything. Marcus was devastated, and trying hard not to show it.
The dramatization is solid — the careful phone placement, the tight jaw, the flattened voice doing the opposite of what he wants. Then the last sentence arrives and tells us exactly what we just watched, as if the narrator doesn't trust us to have felt it. Here's the same paragraph after applying the delete-the-tell strategy with an extended beat:
Marcus set the phone down carefully, like it might go off again if he moved too fast. His jaw was tight, and when he finally spoke, his voice came out flatter than he intended. "They're not coming." Elena watched him for a long moment before she said anything. He picked the phone back up, turned it over once in his hand, and set it down again in the exact same spot.
Nothing is explained. The small repeated gesture — picking the phone up and putting it back exactly where it was — does more emotional work than "devastated" ever could, because it shows a man trying to re-assert control over something small when the big thing is out of his hands. That's the trade you're making every time you cut the tell and extend the show: you lose the guaranteed clarity of the stated emotion, but you gain a texture that a reader actually has to lean into and interpret, which is where the real engagement lives.
If this were a scene with lower emotional stakes — say, a side character getting mildly annoyed in a scene that needs to move along — I'd probably go the other direction and cut the gesture entirely, leaving something closer to "Marcus was clearly annoyed, but he let it go." Not every moment earns three sentences of dramatization. Matching the cut strategy to the scene's importance is the actual skill here, not applying one rule everywhere.
Where This Fits in Your Revision Process
This pass works best after your structural edits are locked and before your line-level polish, since restructuring a scene after you've already smoothed its prose tends to undo the smoothing. If you're following a structured revision order, this slots in neatly with the work described in how to edit a book with AI, and it pairs well with a consistency check — redundant telling often clusters around the same characters or emotional beats, so running an character consistency pass at the same time can surface both issues from the same read-through.
It's also worth running this diagnostic chapter by chapter rather than across the whole manuscript at once. AI context windows can handle long stretches of text, but the comparison task — holding a dramatized beat in mind and checking it against a sentence three lines later — gets less reliable the more text you throw at it in one pass. A chapter at a time keeps the model's attention tight on the actual pairing you're asking it to find. If you want a second layer of automated scanning across the full draft before you go chapter by chapter manually, the Manuscript Cleanup Report catches a lot of adjacent issues — repeated phrasing, pacing drag, and filter-word clusters — that tend to show up in the same paragraphs as these redundant pairs.
One last thing worth saying: this isn't a flaw unique to any one platform or model. It shows up whether you're drafting with a general-purpose chatbot or a tool built specifically for long-form fiction — see Entangled Text vs ChatGPT for more on how different tools handle narrative redundancy. It's a byproduct of how these models are trained to be unambiguous, and it shows up across genres, from romance novel with AI drafts to horror with AI scenes where the dread is supposed to creep, not get announced. Knowing the pattern is half the fix. The other half is just doing the slower read — paragraph against paragraph, not sentence in isolation.
Practical takeaway: run one dedicated pass, chapter by chapter, using a prompt that explicitly asks the AI to find dramatized beats followed by same-sentence or same-paragraph restatements — not generic telling. For each flagged pair, default to deleting the tell and extending the show by one beat, unless the scene is low-stakes and needs speed, or the emotion deserves a delayed payoff later in the chapter. That's three decisions per flagged pair, and all three are faster to make than they are to describe.