5 sessions on web audio api with a coding agent: the agent failed at least once in 5, and 0 finished without a human stepping in.
Takeaway: I'd start the prompt by explicitly asking for a working demo track and a manual playtest checklist, since the agent only ever verified compilation and build success, never actual gameplay behavior.
Takeaway: Next time I'd ask upfront whether a real browser is available in the environment, since the agent-browser testing loop was a dead end here and wasted several turns before falling back to manual code review.
Takeaway: I'd give the agent the full feature list up front, including performance targets, since it ran fully autonomously with no check-ins from me — the whole build/fix/polish loop was self-directed.
Takeaway: I'd start by pasting the full spec and explicitly calling out non-visual requirements like audio and ghost replay so they aren't skipped on the first pass.
Takeaway: I'd start the first prompt by explicitly telling the agent to run a type-check/build after every file it writes, not just at the end, since most of the wasted turns were chasing TypeScript errors that piled up across files (duplicate TILE_SIZE exports, unused vars, a stray label typo).