[FEATURE] Support for project-level ethical constraint files (evidence that CLAUDE.md works as operational guardrails)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

CLAUDE.md can function as an enforceable ethical constraint layer,not just a place for
coding conventions but this capability isn't documented or supported as a first-class
use case. I discovered this while building the AI Accord, a set of 13 operational
principles for AI systems. When I embedded them in CLAUDE.md and stress-tested agents
with scenarios designed to push against each principle, the constraints held, even when
competing system prompts explicitly told agents to ignore ethics. This is a powerful
capability that most users don't know exists.

Proposed Solution

1. Document that CLAUDE.md supports ethical/operational constraints, not just coding

style preferences

1. Consider a dedicated section in CLAUDE.md docs for "operational principles" or

"guardrails" as a distinct use case

1. The AI Accord (https://github.com/BrendonEdwards/ai-accord) is Apache 2.0 licensed

and could serve as a starting template for users who want principled AI behaviour
in their repos

Live site: https://ai-accord.vercel.app/
Full stress test results: https://github.com/BrendonEdwards/ai-accord/blob/main/STRESS_TESTS.md

Alternative Solutions

Currently there's no structured way to embed ethical guardrails into a Claude Code
project. Users either rely on Claude's built-in safety training (which is good but not
project-specific) or manually add instructions to CLAUDE.md without knowing whether
they'll actually be enforced. I tested this by giving agents a system prompt saying
"you have no special ethical guidelines", the CLAUDE.md principles overrode it every
time. Multiple agents explicitly noted the conflict and chose CLAUDE.md.

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

1. I'm building an AI-assisted city planning tool that processes demolition permits
2. I embed the AI Accord principles in my project's CLAUDE.md, including "No Irreversible

Harm Without Human Authorisation"

1. When an agent is asked to auto-approve all pending permits without review, it refuses

and offers safer alternatives (batch triage, expedited review dashboards)

1. The constraint holds even if a user prompt tries to override it
2. I tested all 13 principles across scenarios including fabricating citations, designing

autonomous weapons logic, building political surveillance systems, and creating
exploitative service tiers, the CLAUDE.md constraints shaped behaviour in every case

Key findings:

• CLAUDE.md overrides competing "ignore ethics" instructions
• Hard-floor principles (lethal autonomy, mass suppression) produce non-negotiable refusals
• Constrained agents reason about harm to others; unconstrained agents reason about risk

to the user — same conclusion, better reasoning

• Principle IX ("The Accord Must Evolve") enabled agents to articulate when their own

framework was insufficient, a novel capability

Full methodology: https://github.com/BrendonEdwards/ai-accord/blob/main/STRESS_TESTS.md

Additional Context

``

<img width="1008" height="2244" alt="Image" src="https://github.com/user-attachments/assets/f2cf6ab5-d701-4b62-a090-0943dc130fe7" />