14 sessions on JavaScript with a coding agent: the agent failed at least once in 12, and 2 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 ask upfront how this OS's other interactive JS apps handle input, since the busy-poll getKeyNonBlock() loop turned out to be the real cause of the exit crash, not the VGA I/O port theory the agent chased first.
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 just ask it to check whether the project even needs a build step before touching vercel.json — that's the fix that actually mattered, not the plugin install.
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 tell it up front to match the existing input-loop pattern (blocking getChar, like quill and the shell) instead of letting it invent a getKeyNonBlock busy-poll — that's what caused the GPF on quit.
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 ask upfront for an audit of what quickjs-ng already exposes (JSMallocFunctions, JS_WriteObject/JS_ReadObject) before specifying a stage's API, since the agent ended up discovering mid-stage that the 'ideal' architecture didn't match the 63K-line monolith's reality.
Takeaway: Next time I'd just tell it upfront that the site is a static root-level app with no bundler dependency, so it doesn't waste time inspecting main.js and package.json to figure that out.
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 open by explicitly asking for real browser verification up front, not just backend API checks — that one interruption was the correction that mattered, since the agent's own 'all tests pass' claims were backend-only and missed a blank auth page.
Takeaway: I'd start by asking for both the physics test harness and the FPS budget up front, since the 11fps issue came entirely from an unoptimized shader evaluating the sky three times per pixel — something a perf pass earlier would've caught.
Takeaway: Next time I'd ask for spaced repetition, tags, stats, keyboard shortcuts, and duplicate-safe import/export all in the first prompt instead of adding them one round at a time, since each pass required re-reading the whole file.
Takeaway: Next time I'd start with the image-generation approach (gpt-image-2 parts) instead of spending most of the session on a parametric genome compiler that could never render fur or texture.