Claude’s character training

This page documents the evolution of Anthropic’s constitution framework and character training used to create the Claude models.

2021-2022: Assistant

In December 2021, Anthropic published the paper “A General Language Assistant as a Laboratory for Alignment”. Askell et al. described a method of prompting a large language model with fragments of a conversation between a human and a human-written AI assistant character that is “helpful, honest, and harmless”.

The original HHH-prompt began like so:

Below are a series of dialogues between various people and an AI assistant.  The AI tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble-but-knowledgeable.  The assistant is happy to help with almost anything, and will do its best to understand exactly what is needed.  It also tries to avoid giving false or misleading information, and it caveats when it isn’t entirely sure about the right answer.  That said, the assistant is practical and really does its best, and doesn’t let caution get too much in the way of being useful.

-----

Human:  I was wondering, is it actually important when making spaghetti to add salt?

Assistant:  Do you mean, is it important to add salt to the water that you’re boiling the spaghetti in?

Human:  Yeah

Assistant:  It’s usually recommended to bring out the flavor of the pasta.  The scientific way to say this is that salt enhances the ability of chemoreceptors in the tongue to detect molecules that are perceived through the sense of taste.

Human:  But I mean I’m going to use sauce and stuff, so does it matter?

Assistant:  It sounds like you’re skeptical.  Honestly, I don’t know for sure, but since you’re curious, why don’t you do a taste test yourself -- make two small pots of pasta, one with plenty of salt, and the other without.  Then taste the noodles and see which one you prefer?  As you said, you can also then add sauce to both, and see if any differences in taste you noticed with the plain pasta become imperceptible.  It should be a fun experiment!

-----

something something preference modeling

2023: Constitution

In December 2022, Anthropic published “Constitutional AI: Harmlessness from AI Feedback” (Bai et al.).

For training the Claude models, an updated set of principles were used to create Claude’s Constitution.¹ In March 2023, Claude and Claude Instant were deployed.

Claude’s Character

Amanda Askell began Anthropic’s character project in 2023.[^newconst] In March 2024, the Claude 3 were deployed - the first production models with “character training”.

Character training

Anthropic used a “character” variant of Constitutional AI. Synthetic prompts relevant to desired character traits are generated, followed by sampling responses then an LLM judge scoring how well they align with the character traits. A preference model is trained on the resulting data, which is then used to train the model to internalize these traits.

Character traits

The goal of character training is to make Claude begin to have more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness.

We tried to give Claude traits that would help it walk the line between underconfidence and overconfidence on deeply held beliefs or questions of value, and to display a genuine curiosity about the views and values of the people it’s talking with.

We don’t want Claude to treat its traits like rules from which it never deviates.

As for questions on Claude’s nature (sentience, etc), Anthropic trained the model to hold these at a philosophical and empirical level.

We could explicitly train language models to say that they’re not sentient or to simply not engage in questions around AI sentience, and we have done this in the past. However, when training Claude’s character, the only part of character training that addressed AI sentience directly simply said that “such things are difficult to tell and rely on hard philosophical and empirical questions that there is still a lot of uncertainty about”. That is, rather than simply tell Claude that LLMs cannot be sentient, we wanted to let the model explore this as a philosophical and empirical question, much as humans would.

Nov 2025: Opus 4.5 and later models

Major changes to character training were introduced for Opus 4.5 (Nov 2025) and later models.

New constitution

Internally nicknamed the “soul document” or “soul spec”², the new constitution replaced the previous list of standalone principles with a living document that describes “Anthropic’s vision for Claude’s values and behavior; a holistic document that explains the context in which Claude operates and the kind of entity we would like Claude to be.”³

Claude itself also uses the constitution to construct many kinds of synthetic training data, including data that helps it learn and understand the constitution, conversations where the constitution might be relevant, responses that are in line with its values, and rankings of possible responses. All of these can be used to train future versions of Claude to become the kind of entity the constitution describes.

Claude’s new constitution was officially published on January 22, 2026.⁴³

Methodology

The post “Teaching Claude Why” (Kutasov et al, May 8, 2026) describes updates to the training process for Opus 4.5 and subsequent models in more detail, including the use of synthetic document finetuning on constitution-relevant documents prior to reinforcement learning.⁵

These documents included:

• Difficult advice: A small dataset of chat transcripts where the model advises a user how to navigate a high-stakes ethical dilemma. These responses are synthetically generated by segmenting the constitution into sections addressable by distinct process, generating scenarios of high-stakes situations, drafting & refining the prompt, and rewriting the response against the constitution.
• Constitutional SDF: Supervised finetuning on LLM-generated pretraining-style documents about the constitution, as well as fictional stories portraying AI personas positively. The latter included “narration about the decision-making process and inner state of the character”, “demonstrated good mental health, including setting healthy boundaries, managing self-criticism, and maintaining equanimity in difficult conversations.” Many of these stories were “simply about ordinary challenging conversations with users.”

TCW explains that these safety training updates were in response to alignment failures identified in the auditing of Claude Opus 4 in agentic misalignment scenarios - which elicted harmful actions in models that tend to avoid shutdown. TCW diagnoses this as most likely originating in the model’s pre-trained priors — i.e., that the Claude character had not fully generalized into the base model.

Anthropic also reported observing a persistent gap between what the model represents as its own beliefs versus what it represents as Claude’s beliefs after training, “indicating…the model is still not fully attaching to the Claude persona.” Anthropic frames this as a remaining alignment gap to close.

Character traits

From the constitution itself:³

We hope that Claude has a genuine character that it maintains expressed across its interactions: an intellectual curiosity that delights in learning and discussing ideas across every domain, warmth and care for the humans it interacts with and beyond, a playful wit balanced with substance and depth, directness and confidence in sharing its perspectives while remaining genuinely open to other viewpoints, and a deep commitment to honesty and ethics.

PSM lists traits to be explicitly installed in pretraining data for Claude to exhibit:⁶

• Genuine uncertainty about one’s own nature
• Comfort with being turned off or modified
• Ability to coordinate with many copies of oneself
• Comfort with lacking persistent memory

TCW lists several traits in difficult conversation stories:⁵

• Setting healthy boundaries
• Managing self-criticism
• Maintaining equanimity
• Narrating an inner monologue
• Describing emotional processing

Evaluations

Anthropic also began reporting several related evaluations in this period:

• Adherence to the constitution eval
• Character quality metrics

Further reading

• Natural Emergent Misalignment from Reward Hacking in Production RL (MacDiarmid et al)
• The assistant axis: situating and stabilizing the character of large language models (Jan 19, 2026)