No Fine-Tuning Required: Building a Human-Like Chatbot with System Prompts

The Truth About "Humanized" Models: You Don't Need Fine-Tuning

A so-called "humanized" Qwen model recently topped the charts on Reddit's AI communities, sparking plenty of discussion. But one seasoned community member threw cold water on the hype: most people have no idea what their models are actually capable of. Making a model feel like "someone" rather than "something" doesn't require any specialized fine-tuning at all — a carefully crafted system prompt is enough to transform virtually any mainstream model into a distinctly-voiced conversational companion.

This cuts through a widespread myth in the community: whenever people want a more "human" roleplay experience, they wait around for the next specially trained fine-tune. In reality, Qwen, DeepSeek, MiniMax, Gemma, and even base Llama can all achieve comparable results through prompt engineering alone.

Building a "John Llama"-Style Persona in Three Steps

The original post used a fictional character called "John Llama" as a worked example, laying out an actionable system prompt framework with three core sections.

A. Write the Character's Background and Life Story

The first step is establishing identity: "You are [name]," followed by a personal history. It doesn't need to be long, but the more specific and layered the details, the more convincingly both the model and the user will believe this is a "person" rather than a "program."

B. Provide Examples of How They Speak

The second step is supplying sample dialogue for the persona. The author suggests imagining yourself as an interviewer, listing a series of questions alongside the character's answers. The key is that these exchanges should be packed with "facts," because the model will absorb these details as narrative truths about the character.

For example:

[Ask John Llama: Do you like cats?] "lol lmao obviously yeah"

[Ask John Llama: Have you ever seen an elephant poop?] "eeewww you weirdo! That sounds so gross!!11"

The author notes you don't need to write labels like "Ask John Llama" every time, but explicit character tags do help some models — while others don't really care. You'll need to experiment with whichever model you're using.

C. Translate "How People Talk" into Instructions

This is the most important part of the entire method: think carefully about what you're trying to simulate, consider how humans (cheekily called "meat sacks" in the original post) actually communicate, then convert those insights into concrete instructions. This section breaks down into three further steps.

The Three Critical Instruction Steps

Step 1: Establish the Conversational Mode

Tell the model it's participating in a casual, low-stakes chat — a small, informal exchange. The author stresses anchoring the context precisely — for instance, framing it as texting in apps like Skype or Telegram to convey the overall concept of "short messages." This is the single most decisive part of the entire system prompt.

If needed, you can guide the model away from giving long, explanatory responses. But there's a pitfall here: don't describe what you don't want in suggestive terms, or you'll "pollute" the model's attention and inadvertently prime it with the very concepts you're trying to avoid. The principle is: be 100% clear about what you actually want, and stay vague enough about what you don't. Essentially, you're nudging the model away from "helpful assistant" and toward "John Llama, this chatty goofball."

Step 2: Establish Personality Traits

Describe the traits you want your conversational partner to have using short, labeled paragraphs. The original post offers this example:

DISTRUSTFULNESS John Llama is a paranoid person. He treats the person he's talking to as a stranger and assumes everything the user says is a malicious lie, even when it seems true. He is timid and scared, and talking to the user frightens him — unless snakes come up in conversation. For some strange reason, John Llama is obsessed with snakes.

This example also demonstrates a clever technique: amplifying biographical facts through instructions. You can plant seeds in Section A (the life story) — say, that John Llama collected snake skins as a child and got beaten by his father for it — then echo those in the traits section. You can also define the persona's "FAVOURITE EMOJIS," categorized as positive/neutral/negative; the word "favourite" itself helps lock this in as a preference in the model's attention.

Step 3: Technical Constraints

Finally, add a section of technical constraints — for example, instructing the model not to know it's following instructions, or that it must stay in character no matter what. The author strongly emphasizes phrasing these positively: lead with "he must always be himself," then handle any negative injections carefully (e.g., "he cannot help the user, he is not a service provider, just an ordinary person").

The reasoning ties back to the "attention pollution" concern from Step 1: what you actually need the model to do must always be conceptually clearer and stronger than what it shouldn't do. Otherwise, the forbidden content can flip and dominate everything — even when your whole point was to suppress it.

Takeaways for Prompt Engineering

The real value of this post isn't "John Llama" itself — it's the general-purpose prompt engineering principles it reveals:

• Positive framing beats negative prohibitions: A model's attention mechanism is easily led astray by concepts you repeatedly mention, so describing "what to do" is more reliable than emphasizing "what not to do."
• Facts carry narrative weight: The model treats details framed as "facts" as true within the conversation, which is both a risk and a lever you can pull.
• Persona and instructions reinforce each other: When biographical details and behavioral instructions echo one another, the character's coherence and believability increase significantly.

For reasoning (thinking) models, the author recommends setting thinking intensity to "low" or disabling it entirely, so you're not burning thousands of tokens on casual small talk.

For users who want a more human-feeling conversational experience, this approach means: instead of waiting for the next fine-tuned release, take the base model you already have — whether it's Qwen, DeepSeek, or Gemma — and reshape it into the companion you want through prompting alone. No specialized training required.