What they did
Valentina Koniukhova wanted a trustworthy picture of worklore, an app spread across five repositories: a Python backend on AWS Lambda, a plain JavaScript site, an MCP connector and helper repos. In one day the author ran three open-source Claude Code plugins on it. Floorplan built a clickable map of parts, endpoints and user flows. Archify produced a single architecture diagram with every box tied to lines of code at a pinned commit. Understand Anything, run headless, built a knowledge graph of files, functions and classes with a guided tour. Koniukhova then asked each tool the same three questions about permissions, usage counters and logging, and had a plain Claude Code agent with no map read the code to write an answer key.
How it went
Floorplan first found zero endpoints because the routes live in one table; after the author sent a fix upstream it found 66. Its self-written user flows came from the agent's memory and two were wrong, so the author deleted them and asked Floorplan to trace 17 flows through code; 14 of 16 tested ones worked and the 17th exposed a nightly job that had been failing. Archify took about an hour with two failed runs. Understand Anything took 14 minutes, 16 helper agents and $13.01 at API prices for 732 nodes. None of the three made things up in the quiz, but the plain agent found that OAuth tokens ignored their read and write scopes and could reach the REST API, plus gaps between the privacy page and the code. The author fixed the hole before publishing.
Worth knowing
Ask any generated map a few questions you already know the answers to before trusting it: parts the agent wrote from memory look identical to parts it traced. Archify pins a single repository. Understand Anything's session hook can rebuild the graph without asking if auto-update is on, and its first run burns a lot of tokens. The author runs worklore and links their own write-ups from the post.