How it works
Three steps to
a score you can check.
Pick a challenge, build it with your AI, read the score. The recorder sends events, never your source files.
THE SHIPYARD
Pick a challenge.
Find one in the Shipyard. You might build a feature from a brief, or fix a bug in an app that already exists. Each challenge runs 15-60 min.
Ship a feature from scratch against a spec and a deadline.
Diagnose a broken codebase and land a working patch.
ONE COMMAND
Clone and build.
Paste one command. It clones the challenge repo, connects the GoShip recorder and starts Claude Code. While you work, the recorder logs your prompts, which files changed and the commands you ran.
In Claude Code there is nothing else to install. Cursor and Windsurf take one small config file.
$ claude mcp add goship -- npx -y @goship/mcp-server
GoShip MCP server added
$ claude
Claude Code ready. Start building.
THE SCORING
Get six scores.
The session is scored on AI Fluency, Output Quality, Judgment, Speed, Prompt Precision and Recovery Speed. Published rules score all six. A model also reads a sample of the session and reviews four of them. The methodology page lists the weights.
- AI Fluency
- 92
- Output Quality
- 87
- Judgment
- 78
- Speed
- 95
- Prompt Precision
- 84
- Recovery Speed
- 81
Judgment 78: one large diff accepted without reading it.
PRIVACY
Events, not source files.
The recorder sends events: the prompts you type, which files changed and which commands passed. It never sends file contents, and it skips .env files, keys and credentials. Prompts go up exactly as you typed them, so keep secrets out of them.
Pick a challenge and start.
Your score goes on your profile if you publish it. On a bounty, the winners share the prize.