5 weeks, 43 releases: what kept my Claude Code project (a desktop image editor) from turning into a mess
retained text
I've been building \*\*Scumble\*\*, a free, open-source (GPL-3.0) desktop AI image editor, almost entirely with Claude Code. It started as a ComfyUI node on September 4 and became a desktop app three days later. Today it's on release 43: Electron, a Rust/WebAssembly tile engine for 15,000-pixel pictures, SAM2 and depth maps in the app, and an MCP server with 109 tools, so Claude Desktop or Claude Code can drive the editor too.
I don't write most of the code. I decide what gets built, test it on my own pictures, and say "no, not like that" a lot. A few numbers from the git history: \~83k lines of app JavaScript, about as much test and tooling code, \~43k lines of Markdown plans and docs.
\*\*What actually kept it together:\*\*
\- \*\*CLAUDE.md is the project's memory, not a style guide.\*\* Decisions already made ("do not reopen without the user"), where things stand, a numbered list of everything I've asked for, and a "traps worth keeping" section. Before that list existed, every third session wanted to relitigate Electron vs. something else.
\- \*\*One task per session, then \`/clear\`.\*\* One session that built eight steps in a row pushed the context to 85 %, and quality dropped with it. Now it's one numbered step, its commit, a hand-over note, clear.
\- \*\*Plans with "as built" notes.\*\* Each release has a plan doc; after a step is built, its row says what really happened and what was never tested live. The next session reads that, not its own assumptions.
\- \*\*Say the effect in one sentence before writing code.\*\* I once had a feature fully built before reading a plain description of what it would do to my pictures, and realising I didn't want it. Never again.
\- \*\*Test the real app, by risk.\*\* A gate runner starts a fresh instance with its own profile and drives it over DevTools. I asked for fewer tests at some point, so now there are tiers: full testing only where data can be lost.
\- \*\*Few agents, each with a job.\*\* A builder, a test writer, and a skeptic whose only job is to find what's wrong. Lots of parallel workflows burned my usage limit in an afternoon.
\*\*Favourite traps from that list:\*\* \`globalCompositeOperation = "copy"\` clears the \*whole\* canvas, not your rectangle; Chromium premultiplies alpha on typed-array WebGL uploads too (so pre-premultiplied tiles got darker edges); a 65,536 px canvas silently draws nothing; and \`requestAnimationFrame\` never fires in a hidden window, which hung half our tests.
And one that was nobody's fault but ours: an early benchmark ran on my real profile and overwrote my open document. Every test run gets its own \`--user-data-dir\` since.
The app: [https://github.com/DenRakEiw/scumble](https://github.com/DenRakEiw/scumble) (Windows via the Microsoft Store or installer, Linux builds, macOS coming). Longer write-up with code: DEVTO\_URL
\*\*Question for you:\*\* how do you keep your CLAUDE.md from growing forever? Mine moves old hand-overs to a HISTORY.md, but I'm curious what works for others.document fetched 2026-10-10 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1x1mq0d/5_weeks_43_releases_what_kept_my_claude_code