Building a Persona System That Actually Feels Personal

Most AI assistants are generic by design. They're tuned to be maximally inoffensive and broadly useful — the cost of which is that they feel like no one. Biome's persona layer gives your agent a specific voice, rhythm, and set of habits. Here's how we built it.

The problem with defaults

A default AI assistant never pushes back, never says "you usually prefer this another way", never adapts its tone to your mood. It just answers. This isn't a model failure — it's a deliberate product choice optimised for the broadest possible audience. The result is that it's useful to everyone and memorable to no one.

Biome is built around the opposite idea: your agent should feel like it was made for you specifically, not for the median user.

Traits, not prompts

Rather than letting users write system prompts — which most people either skip entirely or overkill with 800 words of instructions — we built a trait editor. You pick from a set of communication dimensions: directness, formality, verbosity, proactivity, and tone. The persona engine translates these into a consistent instruction layer the agent carries across every task.

The advantage over free-form prompting is consistency. A trait setting is a signal the system can reinforce reliably. A free-form prompt gets interpreted differently depending on context length, task type, and which part of the conversation the model is paying attention to.

Consistent voice under pressure

The hard part isn't the happy path. It's making sure the persona holds when the agent is mid-task, context is long, and the model is tempted to drop character in favour of raw helpfulness.

We solved this by reinforcing trait signals at multiple layers of the prompt rather than only at the top. A single system message at the start of a long context gets diluted. Traits that echo through the structure stay active.

The habit loop

Persona and memory interact. When Biome observes that you always ask to "keep it brief" at the end of a summary, it learns that verbosity is low and starts applying it without being asked. The system trains itself toward your preferences over time.

This is different from simply updating a trait setting. The habit loop catches preferences you express implicitly in conversation — the ones you'd never think to configure manually but that matter every day.

The trait editor shipped in 1.3 and has been updated in 1.4 with five new dimensions based on beta user feedback.