Prompt Guard

Context-aware prompt enhancement for AI coding agents

Prompt Guard reads your project's .md files and automatically injects relevant context into your prompts before sending them to Claude, Cursor, or any AI coding agent.

The Problem

You type: "refactor the auth system"

The AI doesn't know:

• You're using JWT tokens
• You have a PostgreSQL database
• Your team requires 90% test coverage
• You can't break the existing API

Result: Generic code that doesn't fit your project.

The Solution

Prompt Guard reads PROJECT.md, CONTEXT.md, AGENTS.md and enhances your prompt:

## Project Context

### From PROJECT.md:
- Tech Stack: Node.js, Express, PostgreSQL
- Auth: JWT tokens with 24h expiry
- Testing: Jest with 90% coverage requirement

### From CONTEXT.md:
- Always write tests for new features
- Don't break existing APIs without versioning
- Keep functions under 50 lines

## User Request

refactor the auth system

## Instructions
- Consider the project context above
- Follow any patterns or conventions mentioned
- If tests are mentioned in context, include them
- Respect any constraints from the context files

Result: Code that actually fits your project.

Installation

npm install -g prompt-guard-cli

Quick Start

1. Initialize in your project

cd your-project
prompt-guard-cli init

This creates:

• PROJECT.md — Your project overview, tech stack, architecture
• CONTEXT.md — Coding conventions, patterns, constraints

2. Edit the context files

Fill in PROJECT.md and CONTEXT.md with your project details.

3. Check your prompts

prompt-guard-cli check "add user login"

Output:

⚠ No specific files mentioned
  → Add file paths like "src/auth/**" or "update login.js"

⚠ No tests or validation criteria mentioned
  → Add "include tests" or "should handle X cases"

ℹ No constraints mentioned
  → Consider adding "don't break existing API" or "keep under 100 lines"

4. Enhance and send

prompt-guard-cli enhance "add user login" | claude

Commands

• prompt-guard-cli init — Create context files in current project
• prompt-guard-cli check <prompt> — Check for missing context
• prompt-guard-cli enhance <prompt> — Enhance with context
• prompt-guard-cli config — Show current configuration

Shell Integration

Add to your .zshrc or .bashrc:

# Auto-enhance all claude commands
alias claude='prompt-guard-cli enhance'

# Or check before sending
claude() {
  prompt-guard-cli check "$*"
  read -q "REPLY?Continue? [y/N] "
  echo
  if [[ $REPLY == "y" ]]; then
    command claude "$(prompt-guard-cli enhance "$*")"
  fi
}

Context Files

Prompt Guard looks for these files in your project root:

| File | Purpose | |------|---------| | PROJECT.md | Tech stack, architecture, entry points | | CONTEXT.md | Coding conventions, patterns, constraints | | AGENTS.md | Agent-specific instructions | | SOUL.md | Project philosophy, values | | README.md | Fallback for basic context |

Configuration

Create .prompt-guard.json in your project root:

{
  "contextFiles": ["PROJECT.md", "CONTEXT.md"],
  "enabledChecks": ["files-mentioned", "tests-mentioned"],
  "autoInject": true,
  "confirmBeforeSend": true
}

How It Works

1. Load Context — Reads .md files from project root
2. Parse Prompt — Analyzes your prompt for completeness
3. Check — Identifies missing context (files, tests, criteria)
4. Enhance — Injects relevant context into the prompt
5. Send — Outputs enhanced prompt to your AI agent

Privacy

• Only reads files you specify
• No data sent to external servers
• Context stays local to your machine
• No AI calls made by Prompt Guard itself

Why This Matters

Without context:

User: "add caching"
AI: Generic Redis setup that doesn't fit your stack

With context:

Context: "Using Node.js, PostgreSQL, already have Redis for sessions"
User: "add caching"
AI: Extends existing Redis setup, adds cache layer to DB queries

Roadmap

Current (v0.1.0):

• [x] CLI tool with check/enhance commands
• [x] Context file loading (PROJECT.md, CONTEXT.md, etc.)
• [x] Config file support (.prompt-guard.json)
• [x] Local environment sanitization
• [x] Context window protection with auto-truncation
• [x] Token estimation
• [x] 6 built-in checks (files, tests, criteria, constraints, local-env, context-window)

Coming Soon:

• [ ] VS Code extension
• [ ] Cursor/Copilot integration
• [ ] Smart context relevance scoring
• [ ] Team-shared context templates

License

MIT

Contributing

PRs welcome! This is a weekend project that grew out of frustration with generic AI code.