Smart Context Injector for Claude Code

Keyword-aware context injection for Claude Code -- smarter than git status on every prompt.

The common community pattern for Claude Code hooks is to inject git status, git diff, or other project state on every single prompt. That means typing "yes", "ok", or "continue" dumps hundreds of tokens of context that Claude already knows. It wastes your context window, adds noise, and slows things down.

This hook watches what you actually type. It pattern-matches keywords in your prompt against a configurable set of categories, and only injects context when it is relevant. Ask about Django? You get Django context. Ask about Docker? You get DevOps context. Say "yes"? Nothing injected. Zero waste.

User types prompt
       |
       v
 .claude/settings.local.json
 routes to UserPromptSubmit hook
       |
       v
 context-injector.py reads stdin
 (JSON: {"prompt": "..."})
       |
       v
 Load config.json categories
       |
       v
 Keyword matching (case-insensitive)
       |
       |-- No match, or prompt too short --> exit silently (no injection)
       |
       |-- Match one or more categories  --> stdout JSON:
       |                                     {"additionalContext": "..."}
       v
 Claude sees injected context
 as part of the conversation

• Keyword-driven: Only injects context when the prompt matches configured keywords
• Multi-category: A single prompt can trigger multiple categories (e.g., "write a Django test" hits both Python/Django and Testing)
• Short-prompt skip: Prompts under a configurable length (default: 10 characters) are ignored entirely -- "yes", "ok", "go" never trigger injection
• Case-insensitive: Keywords match regardless of capitalization
• Config-driven: All categories live in config.json -- no code changes needed to add, remove, or edit categories
• Zero dependencies: Pure Python 3, no pip packages required

# From your project root:
mkdir -p .claude/hooks
cp context-injector.py .claude/hooks/
cp config.json .claude/hooks/

Replace the example categories with ones that match your project. See Configuration below.

Add the following to .claude/settings.local.json in your project root (create the file if it doesn't exist):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python \"$CLAUDE_PROJECT_DIR/.claude/hooks/context-injector.py\""
          }
        ]
      }
    ]
  }
}

That's it. The hook runs automatically on every prompt, but only injects when keywords match.

All behavior is controlled by config.json, which sits alongside context-injector.py. The format:

{
  "min_prompt_length": 10,
  "categories": [
    {
      "name": "Category Name",
      "keywords": ["keyword1", "keyword2", "keyword3"],
      "context": "CONTEXT [Category]: The text that gets injected when any keyword matches."
    }
  ]
}

• Be specific with keywords. "test" is broad -- it will match "test", "latest", "contest". If that is too noisy, use more specific terms like "pytest", "jest", "test suite".
• Keep context concise. You are paying for these tokens on every matching prompt. State the essentials: where files live, what commands to run, what conventions to follow.
• Prefix with CONTEXT [Label]:. This makes it easy for Claude to distinguish injected context from user instructions.
• Use multiple categories. Five focused categories beat one giant blob. The hook injects only the ones that match.

{
  "name": "REST API",
  "keywords": ["api", "endpoint", "route", "swagger", "openapi", "rest", "graphql"],
  "context": "CONTEXT [API]: API routes are in src/api/routes/. We use OpenAPI 3.1 specs in docs/api/. Authentication is JWT via the Authorization header. Rate limiting is configured in src/middleware/rateLimit.ts."
}

1. Claude Code fires the UserPromptSubmit hook, passing {"prompt": "..."} on stdin.
2. context-injector.py reads the JSON, extracts the prompt text.
3. If the prompt (after stripping whitespace) is shorter than min_prompt_length, the script exits silently -- nothing injected.
4. The prompt is lowercased and checked against each category's keywords (case-insensitive substring match).
5. Every matching category's context string is collected.
6. If any categories matched, the script writes {"additionalContext": "..."} to stdout (categories joined by double newlines).
7. If nothing matched, the script exits silently -- nothing injected.

• Python 3.9+ (uses list[str] type hint syntax)
• Claude Code with hooks support

MIT -- see LICENSE.

Built by VoxCore84