Methodology · CLI · TUI · Community
AI Constitution — Govern your AI assistant with a personal constitution.
Installable. Portable. Auditable.
ai turns the personal AI Constitution
(Constitution.md) into a product.
Install it. Answer a guided interview. Have a personalized
constitution governing every AI tool on your machine — within
thirty minutes.
Install
Get started in one command
brew install convergent-systems-co/tap/aiThen run ai setup — one command installs hooks, wires Claude Code / Copilot / Cursor, installs Claude plugins, and produces your personalized constitution.
Why
Why a personal AI constitution?
- Portable. One markdown file. Every AI tool — Claude Code, Copilot CLI, Cursor, Codex — reads the same governing document. One constitution, all tools.
- Auditable. Every override, every self-noticed violation, every governance decision leaves a permanent typed record under
~/.ai/audit/. Local-only. Never synced raw. - Onboardable. New tools and new AI collaborators read the same files. Onboarding is reading the constitution, not re-negotiating it.
- Self-repairing.
ai doctordetects and fixes the predictable failure modes — broken symlinks, missing hooks, dirty working tree, stale binary — without conversation.
What ships
CLI surface
ai setup, ai review,
ai doctor, ai sync,
ai restore, ai mode,
ai profile, ai persona,
ai skills, ai hooks,
ai settings, ai issue file — every
verb from the spec is on a single binary.
Bubble Tea wizard
A guided interview that produces a personalized
Constitution.md with no fabrication. Every
question offers presets, a free-text answer, or a "chat with
the assistant" handoff.
Atoms ecosystem
Personas, profiles, skills, and brand identity are versioned,
immutable, content-addressable atoms — fetched from four
public registries (persona-atoms.com,
profile-atoms.com, skill-atoms.com,
brand-atoms.com) and cached locally.
Cross-tool enforcement
ai wrap — the cross-platform command dispatcher —
enforces governance at the command layer (git,
gh, opt-in terraform / kubectl)
on macOS, Linux, and Windows. Blocking hooks fail closed; bypass via
absolute path leaves an auditable gap.
Audit by default
Every override, every self-noticed violation, every focus
change leaves a permanent typed record under
~/.ai/audit/. Local-only. Never synced raw.
Self-repairing
ai doctor --fix detects and repairs the predictable
failure modes — missing compact form, stale CLAUDE.md include,
Python App Execution Alias stubs on Windows, missing blocking hooks —
without conversation.
Research
The science behind constitutional AI governance
These four Anthropic papers ground the design of aiConstitution in peer-reviewed research rather than convention.
-
Next-Generation Constitutional Classifiers
Layered principle-based enforcement outperforms reactive pattern matching.
Read on Anthropic → -
Claude's New Constitution
Anthropic's own rationale for why AI needs explicit governing documents.
Read on Anthropic → -
Constitutional Classifiers
Writing the constitution first produces measurably more robust AI behavior.
Read on Anthropic → -
Specific vs. General Principles for Constitutional AI
Hybrid general+specific principles outperforms either approach alone.
Read on Anthropic →
Status
v1.4.4 — production ready
The CLI binary, hook library, and atom registries are live and cross-platform (macOS, Linux, Windows ARM64). The authoritative spec is current at v1.0.0-draft. Source lives at github.com/convergent-systems-co/aiConstitution.