What they did
Michal Kruml had piles of Markdown plans, logs and decision records left over from agent work, spread across a dozen folders, and wanted one desktop app to read them all. He had Claude Code build it using shape-dd, a Claude Code plugin he wrote. The method keeps a short living document with the current decisions, constraints and next unit of work, moves history to a separate log, and runs one unit at a time: a build slice, a decision slice, a re-cut of the plan or a small fix. After each unit the agent stops, he reviews and commits, and the next unit can start in a fresh session. The first unit was not code: the agent checked about twenty older planning documents and reversed five decisions, including moving from Tauri and Rust to Electron, React and TypeScript so he could review the code. A mockup came from Claude Design.
How it went
He reports a working desktop app, MarkScope, after about a week and 65 units, with the full build log public on GitHub. He changed the Settings view after clicking through it, and the agent restructured it in the same session. When he reported only that scroll position was not restored after a restart, the agent inspected the database his test left behind, found saving worked and loading was broken in code from two slices earlier, fixed it, added a test and removed the fix to confirm the test failed. That slice had a manual check that was closed without a real test.
Worth knowing
Stopping after every unit only helps if the person actually does the check they planned. On a longer project he built for days on a data model centred on the wrong thing because every unit looked correct on its own, and his shape file grew past 1,000 lines until he split it. He notes the post itself was drafted with Claude from his logs.