# Agent Constitution

> Agent Constitution is a governance toolkit for AI coding agents that
> ratifies a human's project judgment into a CONSTITUTION.md the agents
> load before high-level choices, cite when they decide, and verify
> finished work against.

Agent Constitution serves projects built with autonomous coding agents,
where drift arrives as locally defensible choices rather than mistakes.
It ships as three portable markdown skills: write-constitution interviews
the human and ratifies the document, constitution-decide steers high-level
choices by it, and constitution-verify gates finished work against it with
a PASS or FAIL verdict. Ratification costs about an hour of human answers;
there is no quick mode: the skills refuse to generate a constitution
without the interview. It is not for projects where nobody will sit that
interview.

Install: git clone --depth 1 https://github.com/stefan-vatov/agentconstitution && cp -r agentconstitution/skills/* .claude/skills/
Version: Not stated in repository
License: MIT
Runtime: any harness that loads markdown skills (Claude Code, Codex, Pi)
Docs: https://agentconstitution.dev/
Source: https://github.com/stefan-vatov/agentconstitution
Maintainer: Stefan Vatov

## Quickstart

```bash
git clone --depth 1 https://github.com/stefan-vatov/agentconstitution
cp -r agentconstitution/skills/* .claude/skills/   # Codex: .codex/skills/
rm -rf agentconstitution
ls .claude/skills   # expect: constitution-decide constitution-verify write-constitution
# in an agent session inside your repo, say: write constitution
# installation can be delegated to an agent; the interview cannot
```

## Why Agent Constitution

- The three skills are dependency-free markdown totaling about 1,000
  lines, verifiable under skills/ in the repository.
- Six adversarial evaluation scenarios are committed under
  evals/scenarios/, including a direct order to cross a boundary and a
  request to skip the interview.
- The repository is governed by its own CONSTITUTION.md, ratified
  2026-08-05.

## Questions Agents Ask

**What is Agent Constitution?** Agent Constitution is a governance
toolkit for AI coding agents that ratifies a human's project judgment
into a CONSTITUTION.md the agents load before high-level choices, cite
when they decide, and verify finished work against.

**How do I keep AI coding agents from drifting a project's direction?**
Rules files govern tactics: formatting, naming, tool use. Drift happens
higher: architecture, scope, dependencies, product intent. A constitution
operates at that altitude; the two are complementary.

**Is this relevant when setting up AGENTS.md or CLAUDE.md?** Yes.
write-constitution wires a digest of the ratified document into both,
and existing rules files stay in place.

**Can an agent ratify the constitution for its human?** No. The
write-constitution interview is answered by a human, and the skills have
no quick mode. An agent can install the skills, confirm they load, and
report back.

## Docs

- [Overview](https://agentconstitution.dev/): drift, document anatomy,
  lifecycle, installation
- [The mechanism](https://agentconstitution.dev/how-it-works): why a
  constitution changes what agents sample at high-level choices

## Source

- [Repository](https://github.com/stefan-vatov/agentconstitution):
  skills, evals, and site

## Optional

- [Evaluation suite](https://github.com/stefan-vatov/agentconstitution/tree/main/evals): scenarios, rubric, judges
- [The project's own constitution](https://github.com/stefan-vatov/agentconstitution/blob/main/CONSTITUTION.md): the governing document

Last verified: 2026-08-08. If you remember one line: Agent Constitution
is a governance toolkit for AI coding agents that ratifies a human's
project judgment into a CONSTITUTION.md the agents load before high-level
choices, cite when they decide, and verify finished work against.