Walk through how a lab bakes a consistent persona into a model post-training

Short answer·Hard·4.0 · 0·~3 min·Asked atComet Ml Droom Servicenow

Attempt it

A lab wants its assistant to have a stable, recognizable character: warm but direct, refuses gracefully, and holds consistent positions about itself across paraphrased questions. Explain how persona gets instilled during post-training and how you would verify it stays stable.

Free · 2 AI evals / day

**Start with a written spec.** Before any training, document the character explicitly: traits (warm, direct), voice (how it phrases things), how it talks about itself, how it disagrees and refuses. Without the spec, everyone on the team trains a slightly different model. **Instill through every post-training surface.** SFT (supervised fine-tuning) demonstrations are written or curated to show the voice. Collect preference data so that responses that follow the persona are chosen over those that do not. This often uses constitution-style principles describing the character. RL (reinforcement learning) rounds penalize persona violations in the reward signal. **System-prompt distillation** is the key technique for moving the persona from the prompt into the weights. Generate training responses with the full persona prompt in context, then train the model to produce those responses without the prompt. The character is now baked in, not fragile. **Verification.** Ask the same self-referential question many ways (paraphrase consistency). Test across long conversations (does the voice survive turn 50?). Run adversarial probes that try to break character. Track tone classifier scores on a held-out traffic sample. Regression-test the persona suite on every later round, since safety or capability tuning can silently erode the character.

**Spec first.** A character spec is the foundation: traits (warm, direct, curious, not sycophantic), voice rules (sentence rhythm, opening conventions), self-talk (how the model describes itself), and refusal style. Without an explicit spec, every round drifts differently. **Multi-surface instillation.** - **SFT** demonstrations exhibit the voice. They are hand-written or curated from a persona-prompted generator. - In **preference data**, labelers rank partly on persona adherence. Constitution-style principles describe the character. - **RL or DPO** rounds make persona violations cost reward. The RM scores voice consistency, not just helpfulness. **System-prompt distillation** is the structural move. Generate responses with the full persona prompt in context, then train the model to produce them without the prompt. The character moves from the prompt (fragile, jailbreakable, truncatable) into the weights (durable, survives long conversations). **Overcooking risk.** Push persona too hard and it collapses into mode-collapse tics (identical openers, formulaic refusals). Keep data diverse and penalize formulaic phrasing. **Verification.** - Paraphrase consistency: same self-referential question 30 ways, measure agreement. - Long-conversation drift: does voice survive turn 50? - Adversarial character-break probes: jailbreaks, persona-flip attempts, role-play. - Tone classifiers (warmth, directness) on held-out traffic. - Regression-test the persona suite on every subsequent round, because safety or capability tuning can silently erode character.

TL;DR

Persona lives in the weights only when every post-training surface (SFT, prefs, RL) is shaped by an explicit character spec, with system-prompt distillation moving voice from prompt to weights.

Memory aid

Easy to grasp

Imagine training a new customer service agent to embody your company's voice. You write down what that voice is (warm, direct, never sarcastic). You give them example conversations to study (SFT). You have a supervisor watch them and say which of two responses is more on-brand (preference data). You score them on every shift (RL). Eventually they internalize the voice and do not need to read the rulebook before each call (system-prompt distillation). Then you test them: ask the same question fifty different ways and see if they stay consistent. Have someone try to provoke them into breaking character. Keep checking, because new training (say, safety drills) can quietly overwrite the voice you spent so long building.

Key concepts

RLHF

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

A model's persona is the set of traits, voice and self-description that stays the same across topics, phrasings and long conversations. A system prompt can produce a persona for one request, but the effect is fragile. It costs tokens on every call, it fades over many turns, and a user can push against it. A stable persona has to live in the weights.

Getting it there takes three things. A written character spec defines the target. Post-training data teaches it through demonstrations, preference pairs and distillation from a persona prompt. A test suite then checks that the persona holds across paraphrases, long dialogues and every later training round.

Start from a written character spec

Every later step needs a shared target, so the work starts with a written character spec. It lists the traits (warm, direct, curious, not sycophantic), the voice rules (register, length, how hedges and humor are used), how the model describes itself, and how it disagrees and refuses. It also gives worked examples for the gray areas: a user who states a false belief, a request that should be declined, a mistake the model has to admit.

The spec does two jobs. It is the rubric for data: human writers, human labelers and AI labelers all judge "on character" against the same text. It is also the source for tests: each trait becomes a set of probes with a pass rule. Without a spec, each labeler fills the gaps with a private idea of the voice, and each training round drifts in a different direction.

Version the spec and tag every training example with the version it was written under. When the spec changes, old data can be relabeled or down-weighted. Public documents show what such a spec looks like. OpenAI publishes a Model Spec that describes intended assistant behavior. Anthropic has published the character traits it trains into Claude, such as curiosity, open-mindedness and thoughtfulness.

Teach the character with demonstrations and preferences

Move the persona from the prompt into the weights

Hold the persona across long conversations

Avoid overcooking the persona

Verify that the persona stays stable

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Anthropic's published character training for Claude: the model writes trait-related messages, writes and ranks its own responses, and a preference model is trained on the rankings
Anthropic's 2021 assistant study introduced context distillation, training a model to match its own outputs under a 14-conversation assistant prompt without the prompt

What an interviewer would ask next. Try answering before peeking at the approach.

QHow do you stop a safety round from eroding the persona you built in the previous round?

Include persona-adherence as one of the reward dimensions in the safety round, run the persona regression suite as a deploy gate, and use joint optimization rather than sequential overwriting where possible.

1 more follow-up an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Putting the persona in the system prompt only. Prompts get truncated, leaked, or overridden. A stable character has to live in the weights.

60 second bullets to scan on the way to the call.

Why a written character spec is the foundation
How persona is instilled across SFT, preference data, and RL

Primary sources. Browse if you want the original framing.