Skip to content
Get packs

Constitutional AI Training

Guides an agent through training a model to self-critique and revise outputs for harmlessness

Source Download skill
Repository
Orchestra-Research/AI-Research-SKILLs
Stars
14,195
Licence
MIT
Files
1 file

What it does

This skill teaches an AI agent the method Anthropic described for making AI models safer without needing humans to label harmful examples. The agent can write example code and a written set of principles (a 'constitution') that a model uses to critique and rewrite its own answers, then train on the improved answers and an AI-judged preference signal. It is a guide for building safety-training pipelines, not a tool that makes a chatbot safe on its own.

Great for

  • ML engineers fine-tuning their own open models for safety
  • Researchers studying AI feedback methods such as RLAIF
  • Teams who want to reduce harmful outputs without a large human labeling effort

Skip it if

  • People who just want a safer chatbot answer: the skill does not change how an agent behaves in a normal chat, and the code is a template that needs a base model, data and compute.
  • Production safety filtering: the skill itself points to runtime guardrail tools such as NeMo Guardrails or LlamaGuard instead, and the AI-judged labels can be noisy or biased.

Try asking

  • “Walk me through the self-critique and revision phase of Constitutional AI for my base model”
  • “Write a constitution with principles that reduce preachy refusals while keeping outputs harmless”
  • “Set up an RLAIF preference-labeling step using my own constitution”
  • “Help me set up a chain-of-thought critique prompt to check responses for helpfulness and honesty”
Step-by-step guide: how to use Constitutional AI TrainingInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 1 original files, the MIT licence and an attribution note.

    constitutional-ai.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip constitutional-ai.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip constitutional-ai.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 3

    Ask for the task

    No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.

Get better results

  • Treat the code as a starting template: the placeholders (base-model, create_dataset, parse_preferences, model_1) must be implemented before it runs.
  • Review the constitution and generated training data yourself, since AI-generated critiques and preferences can carry the judging model's biases.
  • Start with a small dataset and evaluate outputs before scaling up training runs.

What's inside

  • SKILL.md

Redistributed unmodified from Orchestra-Research/AI-Research-SKILLs under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.