npx phantom-dev score

  HARNESS SCORE
  ═══════════════════════════════════════════

  Repo:  my-project
  Score: 34/100 (C)

  █████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░

  AI Configuration          0/20  ░░░░░░░░░░░░░░░
  Code Structure           12/15  ▓▓▓▓▓▓▓▓▓▓▓▓░░░
  Test Infrastructure      10/15  ▓▓▓▓▓▓▓▓▓▓░░░░░
  Naming Consistency       10/10  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
  ...

  TOP IMPROVEMENTS
  → Add a CLAUDE.md with project conventions
  → Add tests for AI to verify against

Quick Start

# Score your repo (just see the results)
npx phantom-dev score

# Score + generate AI config files
npx phantom-dev init

# Output as JSON (for CI/automation)
npx phantom-dev score --json

One command. Zero LLM calls. Zero tokens. 100% free.

What It Does

1. Scans your codebase via static analysis — detects language, framework, test runner, linter, naming conventions, architecture layout, CI pipeline
2. Scores your repo 0-100 across 8 categories of AI-readiness
3. Generates custom configs tailored to YOUR repo's patterns:
   • CLAUDE.md — for Claude Code
   • .cursorrules — for Cursor IDE

Why

Every AI coding tool (Claude Code, Cursor, Copilot, Aider) uses generic prompts for every codebase. Same instructions whether you're writing React or Rust, microservices or monolith.

That's why AI-generated code doesn't match your team's patterns.

Phantom fixes this. It reads your repo and generates configs that teach AI tools YOUR conventions. The result: AI that writes code like a teammate, not a stranger.

Scoring Categories

| Category | Max | What It Measures | |----------|-----|-----------------| | AI Configuration | 20 | CLAUDE.md, .cursorrules existence | | Code Structure | 15 | Directory layout, source organization | | Test Infrastructure | 15 | Tests, test runner, test patterns | | Developer Documentation | 15 | README, CONTRIBUTING, docs/ | | Toolchain | 10 | Linter, formatter, package manager | | Naming Consistency | 10 | Variable and file naming patterns | | CI/CD | 10 | CI configuration, Docker | | Repo Hygiene | 5 | File count, directory cleanliness |

What Gets Generated

CLAUDE.md

Custom config for Claude Code:

• Project overview (language, framework, test runner, toolchain)
• Coding conventions detected from your codebase
• Architecture notes (directory structure, layout pattern)
• Testing guidelines matching your test setup
• Language-specific rules and common mistakes to avoid

.cursorrules

Config for Cursor:

• Key conventions from your codebase
• Architecture context
• Language and framework-specific rules

Supported Languages

TypeScript, JavaScript, Python, Go, Rust, Ruby, PHP, Java, Swift, Dart

Supported Frameworks

Next.js, React, Vue, Angular, Svelte, Remix, Astro, Express, Fastify, Hono, Django, Flask, FastAPI, Rails, Laravel, Gin, Actix, Rocket

The Thesis

The Claude Code source leak (March 2026) revealed that Anthropic's $2.5B product is powered by 512,000 lines of harness code around the model. Memory systems, prompt construction, tool orchestration, context management — the harness does more work than the model.

The model doesn't matter. The harness does.

Phantom extracts this insight into a free tool. Make any AI coding tool work better on your specific codebase. No API keys. No subscriptions. Just a better config.

License

Apache-2.0