Reverse engineer a monorepo into an architecture model with parallel subagents

Hiroki Kondo pointed a Claude Code skill at their own karasu monorepo; 13 parallel subagents produced a 4,453-line model with 301 use cases in about 70 minutes for an estimated $41.78.

Done withClaude CodeinGitHub· Nov 21, 2026

What they did

Hiroki Kondo builds karasu, a text language for drawing system architecture, entirely through Claude Code, and has a reverse-architecture skill that reads a repository and writes a karasu model of it, which they had tried earlier on umami and Dify. This week they pointed it at the karasu repository itself, a monorepo holding a CLI, a library, a VS Code extension, a browser app and a Cloudflare Worker. The session ran 13 subagents in parallel, one per domain, then further steps split out the editor parts and filed a GitHub issue, added references to client-side storage, and made a Japanese copy using Sonnet subagents.

How it went

The model came out at 4,453 lines with 10 services, 17 domains, 301 use cases and 258 entities, and the author did not edit it by hand. Each package of the monorepo was drawn as a service, which the author says split cleanly into domains. The main pass took about 38 minutes and $28.13; the whole job took about 70 minutes and $41.78 across the session and 19 subagents. Six of the 13 subagents independently reported that the language could not point at browser storage, which led to a new notation shipped four days later.

Worth knowing

Costs are the author's estimate from token logs at API list prices, not the billed amount. The parallel domain pass was the longest stage and about three quarters of its cost. Because karasu has almost no database, the agents modelled in-memory structures such as the parser's syntax tree as entities, which is why the entity count is ten times umami's.

Try it

With Claude Code· 2 parts to fill in

Read the repository at [repo path] and write an architecture model of it in [karasu, C4 or my modelling format]. First list the services, treating each deployable package as one, and the domains inside each. Then start one subagent per domain, in parallel, to record its use cases, entities, stores and external calls, citing the files each item comes from. Merge the results into one model, and keep a separate list of anything the format could not express. Stop and show me the summary counts and that list before making any change to the format or the repo.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related