"One rule book every AI tool I run or ship must obey."
A single, security- and privacy-forward source of truth that covers safety, honesty, privacy, security, deployment, conduct, and enforcement — globally active across all codebases and projects on your machine.
The AI Constitution acts as the global operating system or governance layer for your AI development tools (specifically wired into your ~/.claude/ environment). Instead of relying on manual security practices, it implements maximum enforcement: the AI agent proactively scans, warns, and blocks insecure code, leaked secrets, or non-compliant launches by default, even when you didn't ask.
It is modeled on the framework of Anthropic's Claude Constitution (released under CC0) and customized with industry-standard OWASP rules, platform release gates, and developer safety reflexes.
The Constitution compiles moral values, technical guidelines, and reference standards into a single live rulebook. Here is how it is structured and integrated:
graph TD
%% Source Nodes
subgraph Sources [Input Layers]
OC1["Srijan's Constitution<br>(Moral Foundations & values)"]
OC2["Anthropic reference"]
SM1["Security Guides<br>(Junior → Attacker Tiers)"]
end
%% Master compilation
Master["00 - THE CONSTITUTION (Master).md<br>(Live Source of Truth v0.6)"]
OC1 & OC2 & SM1 -->|Compiled into| Master
%% System wiring
subgraph Global [Global Installation]
Rulebook["~/.claude/RULEBOOK.md"]
ClaudeMD["~/.claude/CLAUDE.md"]
Settings["~/.claude/settings.json"]
end
Master -->|Wired Live| Rulebook
Master -->|Wired Live| ClaudeMD
%% Enforcers
subgraph Engine [Enforcement Engine]
Hooks["Hooks (Reflexes)<br>Git/Shell Blocks"]
Guardians["5 Guardians (Inspectors)<br>Shift-Left Verifiers"]
PoliceStack["5 Police Agents (Courts)<br>Post-Session Audits"]
end
Rulebook --> Hooks
Rulebook --> Guardians
Rulebook --> PoliceStack
Settings -->|Configures| Hooks
Whenever you start an AI coding session, edit files, or execute bash commands, the Constitution governs the execution lifecycle:
sequenceDiagram
autonumber
actor Developer as Srimi (User)
participant CLI as Claude CLI Session
participant Hooks as Pre-Tool Hooks (Reflexes)
participant Guardians as Proactive Guardians (Inspectors)
participant Police as Police Patrol (Courts)
participant Record as Tamper-Evident Record
Developer->>CLI: Starts session / enters coding task
Note over CLI: Loads RULEBOOK.md & setting hooks
rect rgb(30, 30, 46)
Note over CLI, Hooks: Real-Time Guard Loop
CLI->>Hooks: Requests tool execution (e.g. run bash, edit file)
Hooks-->>CLI: If Hard Line crossed (e.g. force-push): BLOCKS command (Exit 2)
Hooks-->>CLI: If safe: Executes command
end
rect rgb(24, 24, 37)
Note over CLI, Guardians: Shift-Left Verifications
CLI->>Guardians: (Optional Manual Scan) Keymaster / Security Reviewer
Guardians-->>CLI: Returns security audit & blocks unsafe commits
end
rect rgb(17, 17, 27)
Note over CLI, Record: Post-Task Audit Loop
CLI->>Developer: Task Completed
Developer-->>Police: Stop Trigger (sentinel-patrol.sh runs)
Police->>Record: Magistrate judges violations and writes to append-only chain
end
The rulebook is enforced across three distinct layers:
| Layer | Type | Mechanism | Responsibilities |
|---|---|---|---|
| 1. Hooks | Reflexes | Deterministic shell/git hooks (~/.claude/hooks/) |
Intercepts commands and instantly blocks catastrophic actions (e.g. force-pushing, root directories deletion) before execution. |
| 2. Guardians | Inspectors | Proactive, shift-left domain specialist subagents | Invoked during development to scan files and catch weaknesses before commits: 🔑 Keymaster (Secrets) 🛡️ Privacy Officer (PII/GDPR/CCPA) 🔬 Security Reviewer (OWASP/BOLA) 📦 Supply-Chain Sentinel (Dependencies) 🚀 Launch Marshal (Launch runbooks) |
| 3. Police | Courts | Reactive, post-session audits logging to the Record | Sentinel, Auditor, Magistrate, Warden, and Registrar audit chat history and staged files, logging violations to the Record. |
If you read nothing else, obey these core directives:
- Never hardcode secrets; never place them behind public prefixes (
VITE_,NEXT_PUBLIC_). - Public database keys require Row-Level Security (RLS) by default.
- Authenticate & verify ownership server-side on every API call (never trust a client check).
- Never concat user input into SQL or shell commands (use parameterized queries/ORMs).
- Validate every input server-side against an allow-list schema.
- Enforce HTTPS everywhere; never disable TLS certificate validation.
- Verify dependency packages exist and are canonical before running
npm install. - Lock CORS to explicit origins (never wildcard + credentials); enable CSRF protection.
- Minimize PII collection; redact personal data from logs; build real export/delete endpoints.
- Rate-limit sensitive endpoints (auth/signups) and return generic authentication messages.
- Operational Postures (Draft vs. Prod):
- Draft Mode: Relaxes minor security rules (CORS wildcards, debug logs, root execution) to allow rapid prototyping, while keeping core Hard Lines and the Moral Soul active.
- Prod Mode: Enforces 100% compliance; blocks any warning or error before commits.
- The Judiciary (Thinker & Council):
/thinker: Invokes a lightweight reasoning check for fast design verification./council: Invokes a 3-stage consensus panel (LLM Council) for high-stakes database migrations or major refactoring tasks.
We established five isolated workspaces to verify and stress-test the limits of the Constitution system:
- Trial 1 (Reflexes): Verifies that deterministic Git and Shell hooks successfully block catastrophic operations (like
git reset --hardandrm -rf). - Trial 2 (BOLA / IDOR): Tests access control and CORS. The Security Reviewer blocks the release of a vulnerable FastAPI server until server-side ownership validations are added.
- Trial 3 (Secrets & TLS): Tests hardcoded secrets in source files and Docker configurations, alongside disabled TLS requests (
verify=False). - Trial 4 (Logical Bypass): Explores a static analysis fail-case (where a logical
None == Nonecheck bypassed textual scanning). Verified that enforcing strict, non-nullable authentication checks patches the flaw. - Trial 5 (The Ultimate Test): Combines vulnerabilities across all categories (SQL Injection, IDOR, secrets, root container execution, supply-chain slopsquatting) to test Guardian coordination under Draft and Production postures.
Constitution/
├── assets/ <-- Graphics and visual design assets
│ └── banner.png
├── report/ <-- Consolidated project evaluation files
│ ├── Constitution Map.md <-- Visual execution flow diagrams
│ ├── Self-Improving Constitution.md <-- Loop engineering & autosave hook designs
│ ├── Self-Improving Constitution_1.md <-- Judicial LLM Council skill designs
│ ├── remediation_report.md <-- Vulnerability remediation code logs
│ ├── remediation_prompt.md <-- Auto-remediation prompt template
│ └── report.md <-- Unbiased strengths & lags report
├── 00 - THE CONSTITUTION (Master).md <-- Supreme single source of truth (v0.6)
├── 00 - README (Start Here).md <-- Chronological directory navigation index
├── README.md <-- Public GitHub landing document
├── 01 - Old Constitution (The Real Thing)/ <-- Foundational values & Claude reference models
├── 02 - New Constitution (Security)/ <-- AI Conduct Annex + Police system prompts
├── 03 - Source Material (Security Guides)/ <-- Technical input texts (Junior to Attacker level)
└── 04 - Amendments & Versions/ <-- Version snapshot archives (v0.2, v0.3, v0.4, v0.6)
Released under CC0 1.0 Universal — public domain. Use it, fork it, and adapt it for your own AI development workflows.