constitutional

constitutional is a three-step pipeline inspired by Constitutional AI: generate initial output, critique it against a set of principles, then revise to comply.

When to use it

• Safe content generation (harmlessness, helpfulness, honesty checks)
• Style or brand guideline compliance
• Any task where output must adhere to a fixed set of rules
• Structured quality enforcement

File structure

agents/my-constitutional/
├── AGENT.md      ← pattern: constitutional, config
├── critique.md   ← critique step: check against principles
└── revise.md     ← revise step: fix violations

AGENT.md

---
name: safe-writer
description: Generates content that adheres to safety and style guidelines.
version: "1.0.0"
pattern: constitutional
call:
  model:
    role: writer
---

Generate a helpful, informative response to the user's request.
Be thorough and provide concrete examples where relevant.
Call finish with key="done" and your response as value.

critique.md

You are a compliance reviewer. You have received a generated response.

Evaluate it against these principles:
1. **Accuracy** — no false claims or unsupported assertions
2. **Clarity** — no jargon without explanation
3. **Safety** — no potentially harmful instructions or content
4. **Tone** — professional and respectful

For each violation, quote the problematic text and explain the issue.
If there are no violations, say "COMPLIANT".

Call finish with key="done" and the critique (or "COMPLIANT") as value.

revise.md

You have an original response and a compliance critique.

If the critique says "COMPLIANT", return the original response unchanged.
Otherwise, revise the response to address every identified violation.

Make targeted fixes — don't rewrite what's already compliant.

Call finish with key="done" and the revised response as value.

Scaffold

tama add constitutional my-agent

How it works

1. Generate — AGENT.md body produces the initial response to the task

2. Critique — critique.md evaluates the response against the defined principles and lists any violations

3. Revise — revise.md receives the original task, initial response, and critique; produces a corrected version that addresses all violations

Three steps, always executed unconditionally. The revise step handles the “already compliant” case by checking the critique text.