22 sessions of greenfield projects with a coding agent: the agent failed at least once in 20, and 1 finished without a human stepping in.
Takeaway: Next time I'd start with the 'no-editor, photo-in-animation-out' constraint stated explicitly up front, instead of only surfacing it as a question halfway through the architecture doc — it reshaped the whole approach.
Takeaway: Next time I'd give the exact copy (song, project URLs, email, usernames) up front instead of patching them in one by one at the end — that ate a lot of turns.
Takeaway: Next time I'd specify the memory-arena ownership model (who owns terms from clauses vs.
Takeaway: Next time I'd paste the actual CLAUDE.MD requirements (RemixJS) into the very first prompt instead of assuming the agent read it correctly — that one missed instruction cost the most time, especially once it started patching node_modules to fight Windows file locks.
Takeaway: Next time I'd nail down the QEMU/OVMF/WSL boot pipeline first before layering framebuffer and GRUB experiments on top of it — half the pain was Windows/WSL path translation and stale OVMF vars, not the actual kernel or userland code.
Takeaway: Next time I'd spell out exact screen resolutions and the grid/sidebar split ratio in the very first prompt instead of iterating three times on layout — that back-and-forth cost more turns than any of the actual milestone work.
Takeaway: Next time I'd ask it to run the TypeScript check earlier and more often instead of writing all the frontend files first and fixing a pile of errors at the end.
Takeaway: I'd start the first prompt by explicitly naming the root directory and the include/src split up front, since correcting those after the fact cost two early interruptions.
Takeaway: Next time I'd specify target screen sizes and content lengths up front, since the ABOUT overflow only showed up once I actually looked at the rendered page.
Takeaway: Next time I'd put the milestone list, the GitHub-vs-Codeberg cryptoscam split, and the Python-version target (3.7+) all in the first prompt instead of drip-feeding them — the agent had to backtrack on dataclasses and kw_only fields because 3.7 support came in late.
Takeaway: Next time I'd give the full spec plus real breakpoint screenshots up front instead of discovering overflow and mobile issues one screenshot at a time — that back-and-forth on the projects grid alone took five different layout attempts before landing on the paginated 3-column version.
Takeaway: Next time I'd specify the persistent-arena/variable-registry design up front — the query-vs-clause arena confusion cost the most back-and-forth.
Takeaway: Next time I'd put the actual framework name in the first prompt instead of relying on the agent to read CLAUDE.MD correctly - that Remix mixup cost a lot of time patching node_modules on Windows.
Takeaway: Next time I'd start the prompt by naming the exact Reddit endpoints and noting upfront that reddit.com blocks server-side fetches without browser cookies — that alone would have saved the long detour through proxies, User-Agent spoofing, and CORS services.
Takeaway: Next time I'd start the prompt by stating upfront that I'm on WSL2/openSUSE driving QEMU from Windows, since half the wasted turns came from the agent trying WSL bash/PowerShell bridging that kept mistranslating paths.
Takeaway: Next time I'd specify the exact layout rule (like the 80/20 split) up front instead of letting the agent guess at pixel dimensions per resolution — that's the correction that actually simplified the code.
Takeaway: Next time I'd specify upfront that a .env.example should be verified with a read-back before relying on it, since the 'file created' message wasn't trustworthy on its own.
Takeaway: I'd open with the exact directory layout (include/ for headers, src/ for cpp, root at d:\Yggdrasil) since that's the first thing I had to correct.
Takeaway: Next time I'd point the agent straight at the correct working directory path up front — it burned a lot of turns just figuring out where the project actually lived across Bash and PowerShell.
Takeaway: I'd put the milestone list and the github/codeberg cryptoscam split into the very first prompt instead of adding it after — it saved rework later.