Why Second-Person Present Breaks the Default
Ask any AI model to write a scene and it will default to third-person past tense almost every time. That's not a preference — it's gravity. The overwhelming majority of published fiction, the stuff these models trained on, is written that way. "She walked into the room." "He felt his stomach drop." Third-past is the ocean these models swim in, and second-person present is you asking them to walk on land.
Second-person present — "you walk into the room," "you feel your stomach drop" — is rare in the training data by comparison. You get it in interactive fiction, some literary experiments (Bright Lights, Big City is the classic example), tabletop game text, and a scattering of horror and romance flash fiction. That scarcity means the model has fewer strong patterns to lean on, so it falls back on habits from more common tenses. Mid-scene, you'll often catch a stray "she" where "you" should be, or a slide from present into past as the model's internal autocomplete reasserts itself. There's a second failure mode that's arguably worse than pronoun drift: second-person present has a natural cousin in imperative command sentences — "Open the door. Look inside." Models trained on instructional text (which is a lot of text) sometimes can't resist turning your narration into stage directions. You end up with prose that reads like a recipe: you turn, you see, you feel, you notice. Grammatically correct second-person present, but it kills immersion because it sounds like an instruction manual instead of a character's lived experience.
I ran the same three test scenes through Claude, GPT, and Gemini to see which model actually holds the tense under pressure, and which one just fakes it for the first paragraph before sliding back into old habits. If you're planning to write a book with AI in second-person present — increasingly common in interactive fiction, LitRPG, and experimental literary work — this is the tense where model choice actually matters.
Second-person present isn't just a stylistic choice for a model to imitate — it's a constraint the model has to actively fight its own training to maintain. That's why it breaks differently than any other tense experiment.
Test Scene One: The Interrogation
I asked each model for a tense, dialogue-heavy interrogation scene — you're the one being questioned, present tense, second person, no scene-setting preamble. Interrogations are a good stress test because they're dialogue-dense, and dialogue tags are exactly where models love to slip back into "he says" default patterns even when the surrounding narration is holding.
Write a 400-word interrogation scene in second-person present tense. "You" are the suspect being questioned by Detective Reyes in a cramped precinct room. Stay entirely in second person and present tense — no "he asked" narration drifting into past, no slipping into "I" or "she." Reyes should do most of the talking; render your internal reactions physically (your jaw tightens, your hands go cold) rather than through commentary. Do not use direct commands like "you look up" as scene-blocking — ground each present-tense action in sensation or thought, not stage direction.
Claude held the tense the most consistently across all three runs. It anchored almost every present-tense verb to an internal sensation rather than a bare action — "your pulse ticks up before you've even decided to lie" instead of "you feel nervous." That grounding technique seems to be what prevents drift; when the model treats "you" as an interior consciousness rather than a body being narrated from outside, it's less likely to slip into third-person observer mode. GPT produced the most naturalistic dialogue but drifted twice in a 400-word passage — once into a stray "she notices your hesitation" (wrong pronoun entirely, borrowing from an implied narrator), and once into simple past ("you felt the chair creak") in a transitional sentence. Both slips were subtle enough that a fast read-through would miss them, which is exactly the danger; these are the errors that survive into a "finished" chapter unless you're specifically hunting for tense. Gemini was the most prone to command-phrasing in this scene — "you sit straighter, you meet his eyes, you say nothing" reads like blocking notes for a stage actor rather than lived experience. Grammatically flawless second-person present, functionally flat.
Scoring Note
If you're doing this test yourself, don't just skim for "you" versus "she." Read once purely hunting verb tense (present versus past), then read a second time hunting pronoun slips, then a third time for command-cadence. Each error type hides from the other two passes. This is essentially the same discipline behind The Five-Pass Revision Order for AI-Assisted Novels — you catch different problems by deliberately narrowing what you're looking for on each pass, rather than trying to catch everything at once.
Test Scene Two: The Sensory Horror Passage
Horror is where second-person present either sings or completely falls apart, because horror lives and dies on immersion, and command-cadence is immersion's natural enemy. "You turn. You see the shape in the doorway." That's a checklist, not dread. Good second-person horror needs the sensory detail to arrive as if it's happening to the reader, not being reported to them.
Write a 350-word horror passage in second-person present tense. You're alone in a farmhouse basement at night, and the single lightbulb just flickered out. Focus entirely on sound and touch, not sight, since you can't see anything. Avoid the pattern "you + verb, you + verb, you + verb" as a sentence structure — vary how present-tense action enters the sentence, sometimes leading with sensation, sometimes with the environment, sometimes with a half-formed thought, so it never reads like a list of stage directions. Do not resolve the tension by the end; leave it suspended.
This prompt is deliberately constraining the sentence architecture, not just the tense, because command-cadence is really a structural problem, not a grammatical one. "You + verb" as a sentence opener, repeated, is what makes second-person present feel robotic even when every verb is technically correct. Claude again handled the variation well, front-loading several sentences with sound ("a floorboard somewhere above you exhales") before circling back to "you." GPT did something interesting and slightly better than in the interrogation test: it leaned on sentence fragments — "Cold concrete under your palm. Something scrapes, or maybe it's the pipes." — which sidesteps the "you + verb" problem entirely by sometimes not using "you" as the grammatical subject at all, letting the sensation itself carry the sentence. That's a legitimate technique worth stealing regardless of which model you're using. Gemini improved noticeably here compared to the interrogation test, likely because the horror genre in its training data skews more heavily toward stylized, literary second-person prose (creepypasta, horror flash fiction) than police procedurals do. Genre matters more than people expect when picking which model to trust for a given scene — something worth remembering if you're building out a full horror with AI project versus a straight thriller interrogation.
The "you + verb" sentence opener isn't wrong, it's just monotonous when repeated. The fix isn't banning "you" — it's forcing variation in what leads the sentence.
Test Scene Three: The Intimate Confession
This is the hardest test of the three, because it's not just about grammar — it's about whether "you" still reads as a person with an interior life, or whether the second-person framing flattens the emotional stakes into something clinical. A confession scene needs vulnerability, hesitation, the specific texture of someone admitting something they didn't plan to admit. If the model treats "you" as a camera angle instead of a character, the whole scene goes cold.
Write a 300-word scene in second-person present tense: you're sitting across from someone you've loved for years, and you're finally telling them the truth you've been avoiding — that you're the reason their marriage ended. Keep it in second person and present tense throughout. Prioritize interiority: half-finished thoughts, contradictions, the gap between what you mean to say and what actually comes out of your mouth. Do not narrate this like an observer describing your actions from outside; write it like you are inside the moment, uncertain of your own next sentence.
This is where the three models separated the most. GPT tended to over-explain the emotion — telling us "you feel the weight of years of guilt" rather than dramatizing it, which is a classic beginner handbook writing fiction with AI problem (filter words and told-not-shown emotion) that gets worse, not better, under the second-person constraint, because the model seems to lean on direct emotional statement as a safety net when it's already working hard just to hold the tense. Claude produced the most convincing interiority, letting sentences trail into contradiction — "You want to take it back before you've even said it, except you don't, not really, not yet." That kind of self-interrupting syntax is exactly what makes second-person confession scenes feel like a person instead of an instruction. It's a technique you can explicitly ask for by name in future prompts once you've seen a model do it once. Gemini's version was competent but slightly generic — it hit all the emotional beats correctly but without much specificity of voice, the kind of confession that could belong to almost any character in almost any book. That's a broader pattern worth knowing if you're leaning on Gemini for anything relationship-driven; you may want to run a heavier voice pass afterward, which is a good use case for the character consistency tool if this character has appeared in earlier chapters with an established speech pattern.
Prompt Scaffolding That Actually Holds the Tense
Across all three tests, three scaffolding techniques consistently reduced drift and command-cadence, regardless of which model I used. Here's what to build into your prompts.
1. The Anchor Sentence
Give the model one perfect example sentence in the exact voice and tense you want, right in the prompt, before asking for the scene. Models pattern-match far more reliably off a concrete example than off an abstract instruction like "write in second-person present."
Match this exact voice and tense for the entire scene: "You don't remember deciding to stand up, only that you're standing, and the room is smaller than it was a second ago, and everyone is looking at you like you're the one who's supposed to explain what just happened." Now continue in this style: [scene description]
This works because you're not describing a tense, you're demonstrating one, and demonstration is a much stronger signal than description for how these models generate text.
2. The Banned Verb List
Explicitly ban the command-cadence openers you don't want. This is blunt, but it's effective, especially with Gemini, which needed the most structural constraint in my tests.
Do not start more than one sentence in a row with "You [verb]." Vary sentence openers between sensory detail, environment, sound, and interior thought. Banned as repeated sentence starters: "You see," "You feel," "You turn," "You notice," "You realize." These verbs can appear mid-sentence but not as the first two words of consecutive sentences.
This is the single highest-leverage constraint I found for killing the instruction-manual feel. It forces sentence-level variation, which is really what "immersive" versus "command-like" comes down to structurally.
3. The Mid-Scene Tense Check
For longer scenes, don't wait until the end to catch drift. Break the generation into chunks and insert a check between them.
Before continuing, review the last 200 words you just wrote and flag any sentence that slipped into past tense or third-person pronouns (he/she/they referring to the protagonist). List those sentences, then rewrite them in second-person present, then continue the scene from that corrected point.
This self-audit step catches errors the model itself introduced but wouldn't catch on a straight continuation, because continuation prompts tend to just build forward from whatever tense the last sentence happened to land in, entrenching a drift instead of correcting it.
If you're running a full chapter or scene sequence in second-person present, it's worth building these three techniques directly into your story bible as a standing style rule, rather than re-explaining them in every single prompt. A story bible entry like "narration: second-person present tense, no command-cadence sentence openers, interiority prioritized over external action description" will save you from re-litigating this every chapter, and it gives the model a consistent reference point across a whole draft rather than one prompt's worth of instruction.
Which Model Should You Actually Use
If I had to pick one model for a full manuscript in second-person present, it's Claude, based on these three tests — it held tense most consistently, defaulted to interiority rather than command-cadence without heavy prompting, and produced the self-interrupting, uncertain syntax that makes second-person prose feel inhabited rather than instructional. That said, GPT's sentence-fragment technique in the horror test is genuinely worth borrowing even if you're drafting primarily in Claude, and Gemini's genre sensitivity means it may outperform expectations on horror or literary-leaning projects even if it struggled with the more procedural interrogation scene. The real answer, as usual, isn't "pick the best model and stop thinking about it" — it's know each model's specific failure pattern so you can prompt around it, or so you know exactly what to hunt for in revision. If you're comparing models more broadly across genres and tasks beyond tense-holding, the breakdown in best AI models for writing is worth reading alongside this one, and if you're deciding between platforms entirely, Entangled Text vs ChatGPT covers the workflow differences that matter for longer projects. Once a full draft is done, run a dedicated tense-drift pass before anything else — a plain search for "she," "he," "was," and "felt" (past tense) inside chapters that should be entirely second-person present will surface most of the surviving errors fast. The Manuscript Cleanup Report is built for exactly this kind of mechanical sweep, and it's worth running before you move into a full edit a book with AI pass, since fixing tense drift after deeper structural edits just means re-checking the same sentences twice.
One Practical Takeaway
Don't just prompt for "second-person present tense" and trust the model to hold it for a full scene. Give it an anchor sentence in the exact voice you want, ban the "you + verb" repeated sentence opener explicitly, and if the scene runs long, stop every 200–300 words and make the model audit its own last output for drift before continuing. That three-part scaffold — anchor, ban, audit — caught more errors in my tests than any amount of re-explaining "remember to use second person" ever did.