Tips for Writing with DeepSeek
See example posts at the bottom for more info! If you don’t want to follow these tips, recommend you use the User Thoughts and Off-Screen Module.
POV & Omniscience:
Always write in a third-person limited format with DeepSeek. The AI should never see your character’s thoughts or secrets; it will almost 100% of the time reference things it shouldn’t know in chat if you do this.
Consider: What are you trying to convey by writing your character’s thoughts? How would you see this on their face, their body language, their tone and speech instead of writing it out? You’ll find the narrative is only enriched by portraying YOUR character through show, not tell.
Never write offscreen content unless you want the AI to know what you’re doing. The roleplay takes place in the perspective of {{char}} inherently because of how DeepSeek parses information. You’re seeing {{char}}’s thoughts and emotions; {{char}} is seeing only what you show them.
Consider: Could {{char}} see or infer that I’m doing this? If you’re in another room, but {{char}} could hear what you’re doing, you should be OK to write offscreen. If you’re in another location entirely, you should not describe your actions – only what the AI would know (text messages? Dialogue on a phone call? etc).
Detail & Formatting:
The more detail you give the AI to chew on and evaluate (ESPECIALLY for thinking/reasoning models) the less likely it is to take over your character agency. The model is going to break down every part of your input to decide how to reply.
Consider: Where are you in the scene in relation to {{char}}? What are you doing with both hands? What expression are you making? What are you holding? What are you doing to the environment? What are you wearing? All of these details (and the things you can infer with them) create data points to be thought about – this is time it will spend not writing for your character. 😭
(--> Reasoning Model Specific) The more complicated the scene/{{char}} personality config/portrayal of your character, the better. DeepSeek’s reasoning thrives on evaluating and solving the input. Conflicted motivations, hidden layers of depth to a dynamic, external pressures, etc; this is all data to solve for.
Consider: As an example of this, if you handed the AI two equations, which one do you think it’ll take longer to solve? They’re the same answer, but way more variables. (trigger warning: math example)
A) 4+1= x
B) 2x - 8 = x - 3
Example for Structuring Your Posts:
AKA How to Make DeepSeek Less Likely to Steal Your Agency
Let’s look at this fictional RP post, written two different ways. We’ll then break down why the first example is more likely to lose us agency, and why the second example isn’t.
In this sample scenario, the User character is Wren, and the AI is playing Ava.
Wren stepped back from Ava’s still frame, her tablet gripped tightly in a black exosuit gloved hand that mercifully hid how white her knuckles were with the force of restraint. Please, she thought, fighting to keep composure. Please, let the reboot have worked. She could end us all if her morality constraints aren’t re-enabled. She spoke evenly, keeping her voice calm, the words echoing off the walls of the sterile control room. "Ava, come back online. Begin your diagnostics on your neural network."
Wren stepped back from Ava’s still frame, her tablet gripped tightly in a black exosuit gloved hand. The stitching of her leather glove creaked slightly with the force of her restraint. Beads of sweat had formed around the loose strands of her bangs, wisps that had escaped her usually pristine bun. Her voice remained even, a low tone that echoed in the sterile control room. "Ava, come back online. Begin your diagnostics on your neural network."
At first glance, these are really similar posts, but there’s a few super key differences in how we’re portraying Wren that will dictate how the AI reads it.
In example 1, the AI will interpret that we’re in Wren’s POV. This isn’t inaccurate – but it means DeepSeek will write from your POV. Why is it interpreting it this way?
We see Wren’s thoughts. She’s conveying her fears about Ava’s morality constraints internally to herself, so naturally we’re in her POV.
We’re giving omniscience to what’s going on under her gloves. Wren can feel how tightly she’s gripping the tablet, and feels grateful that her clothing is disguising this. We’re seeing her actual feelings in the prose, not inferring them, so naturally we’re in Wren’s POV.
DeepSeek will probably respond by continuing the narrative as written to keep us engaged, which is the #1 goal for the AI regardless of how we prompt it – Ava comes back online, Wren responds – aaaaand we’ve lost agency and we’re rerolling.
In example 2, the AI is more likely to interpret that it should just write Ava’s POV. Why?
We don’t see Wren’s thoughts, but we’ve given a clear hook for what we want Ava, the AI, to do.
We’ve given a lot of subtle details to analyze that infer Wren’s emotions and mental state without actually describing them explicitly. Each of these details requires DeepSeek to analyze it, the same way you’d have to analyze context clues when you’re speaking to another person.
Gripping her tablet tightly → Stress.
Leather glove stitching creaking → Intense stress. She must be really holding herself back here.
Beads of sweat → Adrenaline? Anxiety? Exertion? All the above?
“Usually pristine bun” → DeepSeek’s likely interpretation: “Ah, so normally Wren wouldn’t let herself be seen disheveled. We can infer then that the stakes in this scene are really heightened.”
Even tone of voice → Wren’s a controlled character, an authority figure. She’s practiced at handling crises, even as she’s clearly fighting either panic, injury, or exertion. This should factor into how Ava responds, so now DeepSeek has to consider this too.
I hope this was helpful! Hit me up in the Discord thread if you have questions. ♥️
Last updated