Takeaway: Next time I'd start by pointing the agent at real Pauperwave labeled data before letting it guess archetypes from card frequency — that guesswork produced a whole feature I had to mark 'keep open' until we redid it properly.
Takeaway: I'd start the next session by pasting in the actual SOW questions up front instead of doing three rounds of back-and-forth, since most of the ambiguity was around color computation rules and navigation UX that I could have specified from the start.
Takeaway: Next time I'd point the agent straight at .github/workflows/run-crawler.yml and entrypoint.py instead of letting it search broadly, since .cursorignore blocked the workflow file and cost a lot of exploration turns.
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 ask directly for a Node-version-compatibility check before installing a new global CLI, since the crash turned out to be an ESM/extensionless-bin packaging issue rather than an install failure.
Takeaway: Next time I'd start the prompt by asking the agent to first diff the 2026 data against 2025 before touching anything, since the real problem turned out to be a stale scrape, not a missing setup.
Takeaway: I'd start by just asking for the JSONL transcript file path directly since that's really what I needed, instead of having the agent re-derive it through dozens of greps and doc fetches.
Takeaway: Next time I'd start by asking for the HF-native agent-traces format directly instead of discovering the hard way that Cursor's raw JSONL isn't supported by the viewer — that caused a full re-upload of all 14 conversations.
Takeaway: Next time I'd include a screenshot of the target style upfront instead of describing it after the fact — the first redesign attempt with the fancy SVG icon and color-mix CSS looked good in theory but broke visually in the browser.
Takeaway: Next time I'd front-load the dataset ID and scope questions in the first prompt so the agent doesn't have to burn a dozen turns guessing at what I meant by 'seed_dict' vs full retrain.
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.
Takeaway: Next time I'd specify upfront that the agent should stay in Agent/edit mode so we don't lose a turn just switching modes after I approved the plan.
Takeaway: I'd start by pointing the agent straight at the load/init logic (index.html's tournament selection and index.json sorting) instead of letting it explore the whole file top to bottom, since that's where the actual bug lived.
Takeaway: I'd ask for the same thing again but mention up front that init() should fully reset cached state, since that's exactly what needed fixing along the way.