A document review agent in Claude Code

A complete build of the capstone. A register tool and a Zod schema with four code checks feed a gated claude -p chain with capped retries, run by a subagent with a preloaded skill, contained by permission rules and an audit hook, and measured by a 10-case eval before and after one fix.

Back to the project

The system is a chain: run.ts asks a title-reviewer subagent (running under claude -p on your subscription) for one TitleReport per title, then a code gate checks that report against the parsed register and either accepts it, sends the exact errors back for a retry, or hands the title to a human. The decision that matters most is that the model reads and code decides what is valid: the agent's register tool and the gate share one parser, so every quote is checked against the same text the agent was shown. Every document in pack/ is synthetic; the people, lenders, addresses and title numbers (ZZ100001 to ZZ100010) are invented.

Back to the project

A document review agent in Claude Code

A document review agent in Claude Code

Sign in to keep reading the answer.

Sign in to keep reading the answer.