8 sessions on github pages with a coding agent: the agent failed at least once in 4, and 4 finished without a human stepping in.
Takeaway: I'd start the next session by naming the full scope up front — scraping, pricing, registration backend, and payments — since each one kept surfacing new hidden requirements the agent had to dig for.
Takeaway: Next time I'd ask upfront to check the scrape schedule against the actual tournament dates, since the day-1-only cron was the root cause of missing results.
Takeaway: Next time I'd lead with 'the malus is making everyone's score negative during the group stage, only show earned points until knockouts' since that was the change that actually mattered to users watching the live tournament.
Takeaway: Next time I'd mention upfront that .cursorignore blocks the repo root, so the agent doesn't waste turns rediscovering that and has to fall back to fetching the public GitHub mirror.
Takeaway: Next time I'd paste the exact regolamento rule text into the first prompt instead of describing it from memory, since the agent's first read of rule 7 was wrong and I had to correct it after it built and tested the opposite behavior.
Takeaway: Next time I'd open straight with 'fix where colors are assigned in the CI pipeline, not just the local script' — that's what saved a round trip, since the bug was in the CI-only entrypoint.py, not the newer pipeline.py I'd have guessed first.
Takeaway: Next time I'd mention upfront that .cursorignore hides root-level files, so the agent doesn't have to discover that on its own and fall back to GitHub API calls.
Takeaway: I'd start next time by asking directly for the crawling pipeline simplification instead of first exploring the general architecture, since that scoping question and plan review was where the real value came from.