14 sessions on onnx with a coding agent: the agent failed at least once in 8, and 5 finished without a human stepping in.
Takeaway: Next time I'd tell it upfront that the official docs/install links might be stale or gated, since chasing the broken GitHub skill repo and rebuilding the skill from scratch ate a huge chunk of the session.
Takeaway: This wasn't really a build — I used the assistant purely as a research reference to survey datasets, models, fine-tuning approaches, and deployment options before deciding on my own stack.
Takeaway: Next time I'd scope this to one decision at a time instead of firing off 260+ open-ended questions in one sitting — I never actually got to building anything, just collecting options.
Takeaway: I never actually pointed this at a real task — I just fired off question after question and never gave the agent a concrete file or repo to work with, so it kept guessing at a 'Next.js project' it couldn't inspect.
Takeaway: I used this as a rapid-fire research session rather than a coding task, so there was nothing to intervene on or fix — it was just me pulling information across a huge surface area.
Takeaway: Looking back, this wasn't a build at all — it was me firing off dozens of unrelated research questions across the ML stack, and the assistant answering each in isolation with no continuity or actual implementation.
Takeaway: I'd frame this upfront as a single research brief instead of firing off 250+ standalone questions, since the assistant kept resetting context and repeating boilerplate advice rather than building on prior answers.
Takeaway: Next time I'd specify upfront that the model runs client-side so we should keep full semantic context (all synsets, full glosses) even under credit/perf constraints — the agent's instinct to shrink things for a WebGPU error nearly cost us disambiguation quality.
Takeaway: Next time I'd mention upfront that streaming/shuffling a chronologically-shaped HF dataset is unreliable for randomness, and just point straight at the local summary.json/transcripts.csv files I already had instead of letting the agent try three different remote sampling strategies first.
Takeaway: Next time I'd ask upfront how the HF dataset is sharded/ordered so we don't burn a whole session discovering that streaming shuffle only randomizes within a season-grouped window.
Takeaway: Next time I'd have it check for a working Python runtime and gradio version up front, since that cost several minutes of failed docs searches and PATH issues before it could even inspect the API.
Takeaway: Next time I'd tell it upfront not to bother with in-app browser or Playwright automation on this machine — just get the server running and let me check it myself in a real browser.
Takeaway: I'd specify upfront that mobile-responsiveness testing should rely on manual review or a known-working tool rather than have the agent hunt for a broken local Playwright install, since that ate time for no result.
Takeaway: Next time I'd mention upfront that local browser automation tooling (Playwright etc.) is flaky in this sandboxed Windows environment, so we don't burn a long stretch discovering that mid-session.