What it does

Project Ledger is a human-in-the-loop findings system for Claude Code. It provides a structured YAML ledger for logging insights, a /ledger skill for publishing entries (with an optional Haiku auditor), and a UserPromptSubmit hook that runs TF-IDF search on every prompt to inject relevant entries automatically.

What problem it solves

AI agents can't judge what matters — they treat implementation details the same as architectural flaws. Project Ledger lets the human decide what gets remembered and how it's framed, then surfaces those findings automatically via context injection.

Who uses this workflow

Developers working on research, hardware, data, or creative projects where hard-won insights accumulate over time — anyone who wants control over how architectural decisions are tracked and a memory system simple enough to actually use.

How it works

1. Use /ledger to publish a finding — the skill writes structured YAML and optionally spawns a Haiku auditor
2. A UserPromptSubmit hook runs TF-IDF cosine similarity on every prompt
3. Top matching entries get injected as context before the agent starts thinking