The problem with asking a model to mark its own test
Ask any model to draft a scene, then ask it to critique that same scene, and watch what happens. It'll find a few surface issues — maybe a clunky sentence, maybe a weak verb — and then tell you the emotional arc is strong, the pacing works, and the dialogue feels natural. It's not lying to you. It genuinely can't see its own blind spots, because the critique is being generated by the same statistical habits that produced the draft in the first place. This is the same reason you can't proofread your own wedding speech at 2am and catch the typo you've read fifteen times. The model that wrote "her eyes sparkled with unshed tears" doesn't flag that phrase as a cliché, because to that model, it's not a cliché — it's the natural next-token choice, the thing that felt right when it generated the sentence. Self-praise bias isn't a quirk you can prompt away with "be more critical this time." It's structural. The drafter and the critic share a brain, and a brain doesn't audit its own tics. Every model has tics. One overuses em-dashes and sentence fragments for emphasis. Another leans on a small rotation of verbs — "settled," "flickered," "lingered" — until every scene feels like it's happening underwater. Some models resolve tension too fast because they're optimizing for a satisfying-sounding paragraph rather than a slow-burn scene. If you're using that same model to evaluate its own work, those tics are invisible by definition. They're the water it swims in.
A model can't critique what it doesn't know it's doing. You need a second set of eyes that owes the first model nothing.That's the whole premise of the three-model relay: separate the writing from the judging from the fixing, so no single model gets to grade its own homework.
Assigning roles: the Drafter, the Critic, and the Rewriter
The relay works because each role has a different job and, ideally, a different brain doing it. The Drafter writes the scene. This is the model you'd normally use to write a book with AI — whatever you've tuned your voice prompts for, whatever's been tracking your story bible and character sheets. It doesn't need to be defensive about its own prose because it's not going to be asked to judge it. The Critic reads the scene cold, with no investment in how it turned out, and its only job is to find problems. Not fix them — find them. This model should ideally come from a different family entirely. If your Drafter is Claude, make your Critic GPT or Gemini. Different training data, different RLHF tuning, different default instincts about what "good prose" means. That divergence is the entire value. A critic trained on the same patterns as the drafter will miss the same things. The Rewriter takes the critique — not the original draft's self-assessment, just the critique — and produces a new version. This can be the original Drafter model again, but it has to be a clean context. No memory of writing the scene, no attachment to that turn of phrase it liked. Fresh eyes, prior instructions only. If you're working inside Entangled Text, this is easy to set up because you can run different models against the same manuscript context without juggling browser tabs and copy-paste — see the Entangled Text documentation for how model switching works per-chapter. If you're stitching this together yourself across ChatGPT, Claude, and Gemini tabs, it's more manual but the principle holds: three distinct contexts, three distinct jobs. Check out our comparison of Entangled Text vs ChatGPT if you're deciding whether to consolidate this workflow into one tool. One thing worth flagging early: this relay is a diagnostic layer, not a replacement for structural revision. It catches sentence-level and scene-level problems. For the bigger developmental issues — pacing across chapters, arc consistency — you still want something like The Five-Pass Revision Order for AI-Assisted Novels running alongside it.
Prompting the Critic to find problems, not fix them
This is where most people blow the whole exercise. They ask the Critic to "give feedback" and the Critic, being a helpful assistant by training, immediately starts rewriting lines inside its own critique. "This sentence is weak — try: 'The rain fell harder.'" The second it does that, you've lost the separation. Now the Rewriter is just polishing the Critic's prose instead of solving the actual problem, and you're back to one model's taste dominating the output. You have to explicitly forbid rewriting in the prompt. Tell the Critic its job ends at diagnosis. No suggested replacement lines, no "instead try." Just: here's what's wrong, here's why it's wrong, here's where it happens.
You are a developmal editor reviewing a scene for a client. You have no attachment to this prose — you didn't write it and you don't need to be kind about it. Your only job is diagnosis, not repair.
Read the scene below and identify every instance of: (1) telling emotion instead of showing it through action or dialogue, (2) repeated sentence structures or rhythms that create monotony, (3) dialogue that sounds like exposition dressed up as conversation, (4) any moment where tension is resolved faster than it should be, (5) clichéd phrasing that a careful human editor would flag.
For each issue, quote the exact sentence or passage, name the problem category, and explain in one sentence why it weakens the scene. Do NOT suggest a rewrite, alternate phrasing, or fix of any kind. If you catch yourself writing a replacement sentence, delete it and just describe the problem instead.
Scene: [paste scene]
This works because it gives the Critic a checklist instead of an open-ended "what do you think," which forces it past the lazy "this is good, nice job" response. Naming specific categories — telling vs. showing, repeated rhythm, exposition-as-dialogue — gives it something concrete to hunt for instead of vibing a general reaction. Swap the five categories depending on genre: if you're troubleshooting a romance scene, you might ask it to check against the beats in our Romance Beat Sheet: Where AI Drafts Usually Break; for a mystery, you might ask it to flag clues that are either too obvious or dropped without payoff, which pairs well with the instincts in our mystery novel with AI guide. A second prompt, useful once you've got the categorized list and want the Critic to rank severity:
Here is your previous list of issues with this scene. Rank them from most damaging to least damaging to the reader's experience, where "most damaging" means a reader would notice and be pulled out of the story, and "least damaging" means only a careful re-read would catch it.
For the top three issues, explain specifically what effect they have on pacing, tension, or character believability. Do not propose solutions — only explain the damage.
This stops you from treating every nitpick as equally urgent. A repeated sentence structure three paragraphs apart is not the same severity as a character acting against their established motivation — but an uncalibrated critique list will present both with the same flat tone. Ranking forces the Critic to actually weigh the problems against each other, which is closer to what a real editor does during how to edit a book with AI passes.
Feeding critique into the Rewriter without the original draft biasing the fix
Here's the counterintuitive part: don't give the Rewriter the original draft and the critique side by side. If you do, the Rewriter tends to anchor on the existing sentences and make minimal, surface-level edits around them — swap a word here, tweak a clause there — because the original prose is sitting right there in context, exerting gravitational pull. You end up with a patched version of the same scene, tics and all, instead of a genuinely reconsidered one. Instead, try giving the Rewriter only the critique and the surrounding story context — the story bible info, character voice notes, what needs to happen plot-wise — and let it write the scene fresh, informed by what to avoid rather than what to patch. This is closer to how a ghostwriter would work from an editor's notes than how a copyeditor works with tracked changes.
You're drafting a scene for a novel. You have NOT seen any previous draft of this scene — write it fresh. Below is (1) the story context you need and (2) a list of specific pitfalls identified by an editor reviewing an earlier attempt at this scene. Your job is to write a version that avoids every pitfall listed, without reference to how the previous draft handled it.
STORY CONTEXT: [character voice notes, what must happen in the scene, where it falls in the plot, POV character]
PITFALLS TO AVOID (from editorial review of a prior draft):
1. Telling emotion instead of showing it — specifically, naming feelings directly ("she felt a surge of anger") rather than letting action or dialogue carry it
2. Three consecutive sentences in the previous draft started with "She" — vary sentence openings and lengths
3. The confrontation was resolved in one exchange; the previous version had the antagonist back down too quickly given his established stubbornness
4. Dialogue in the previous draft explained backstory the reader already knows — avoid any dialogue that functions as a recap
Write the scene now, roughly 800-1000 words.
Why this works: the Rewriter is building from intent rather than editing existing text, which means it's not fighting the gravitational pull of sentences it didn't choose. It also forces you, the author, to translate the Critic's diagnosis into actual constraints — which is a useful discipline on its own, because it stops you from passing along vague feedback like "make it punchier" that no model can act on cleanly. If you're worried about continuity — did the Rewriter keep the character's established voice, did a detail from three chapters back get dropped — run the new draft through a tool like the AI character consistency checker afterward, since a from-scratch rewrite is slightly more likely to drift from established voice than an edit-in-place would be.
When to stop the relay
The relay has diminishing returns, and if you don't know where to stop, you'll over-correct a scene into blandness. I've seen this happen a lot: round one fixes real problems, round two fixes smaller problems, round three starts inventing problems because the Critic has been told to find five issues and will find five issues whether or not they're actually there. Watch for these signs that you've gone far enough:
- The Critic's issues are getting smaller and more subjective each round — round one flagged "the antagonist's motivation doesn't track," round three is flagging "this sentence could maybe be shorter." That's not a real problem, that's a model fulfilling a quota.
- The voice is drifting from your established character sheets. Every rewrite pass is a chance for the Rewriter to smooth out the specific roughness that made a character sound like themselves. If your short-tempered character starts reading as merely brisk, you've sanded off a personality trait in the name of "clean prose."
- You read the newest version and genuinely can't tell if it's better than two rounds ago, just different. That's the clearest sign — if you have to think hard about which draft is stronger, stop and pick the one that still sounds like your book.
Two rounds is usually the sweet spot: one draft, one critique, one rewrite. A third pass is worth it only if the second-round critique surfaced something structural that the first missed entirely — a plot hole, a motivation gap, something that genuinely needed another look rather than another polish. If you're running this relay on an opening chapter, pay extra attention to that first round, since first pages carry more weight than any other scene in the book — our guide on opening hooks for AI drafts covers what's actually worth fixing there versus what's just nervous over-editing. It also helps to budget for this upfront rather than discovering three rounds in that you've burned through more tokens than planned — see How to Budget AI Drafting for a Full Novel if you're running this relay across dozens of scenes and want to keep API costs predictable.
Polish removes what's wrong. Over-correction removes what's distinctive. The relay is only working if it's doing the first thing.Once you've settled on a final version, this is a good moment to run a Manuscript Diff between your original draft and the relay's output. Seeing the actual deltas side by side — not just trusting that it "feels different" — makes it much easier to catch cases where a rewrite accidentally dropped a plot detail or softened a character beat that needed to stay sharp.
Making this part of your regular process
You don't need to run every scene through a three-model relay — that's exhausting and unnecessary for scenes that are already working. Save it for the scenes that matter most: chapter openings, major confrontations, anything your beta readers flagged as flat. If you're running a structured Beta Reader Workflow for AI-Assisted Manuscripts, the relay is a natural step to run before you send pages out, catching the self-reinforcing tics a single model would wave through. It's also genre-flexible. A LitRPG writing project has different Critic checklists than a horror with AI project — stat-block consistency and progression logic versus dread pacing and the restraint of what's shown versus implied. The categories you ask the Critic to check should always be genre-specific, not generic "is this good writing" prompts. The actual takeaway: next time a scene feels off but you can't articulate why, don't ask the same model that wrote it to explain what's wrong. Open a clean context, paste in the scene with the "diagnose, don't fix" prompt above, and see what a model with zero investment in those sentences actually finds. You'll usually be surprised by how much was invisible to you simply because you — like the model — already knew what you meant to say.