7 sessions on yolo with a coding agent: the agent failed at least once in 5, and 2 finished without a human stepping in.
Takeaway: Next time I'd ask upfront for a push script that uses create_commit/CommitOperationAdd from the start, since upload_file and upload_folder both got silently skipped by HF's LFS dedup and wasted two rounds.
Takeaway: Next time I'd flag the Windows-specific gotchas (WeasyPrint's native deps, SQLite file locks, libGL missing on Modal's slim image) up front in the first prompt since they cost real cycles to debug reactively.
Takeaway: Next time I'd paste the full Unsloth/Modal error and ask for the actual root cause up front instead of letting quick patches loop on the same symptom.
Takeaway: Next time I'd ask it to actually dry-run or at least trace the Modal image build and dataset serialization steps against Unsloth's current examples, not just review the code — the static review missed both the apt-get ordering bug and the PIL/Arrow serialization issue, which only showed up once I actually ran it on Modal.
Takeaway: Next time I'd ask upfront for the full dataset provenance and org/token permissions before starting a training run, since I ended up chasing HF org migration details mid-run.
Takeaway: Next time I'd paste the actual HF dataset_info metadata into the first prompt so the card is right the first time instead of needing a follow-up fix.
Takeaway: Next time I'd paste the file path and both exact values upfront so the agent doesn't need to read the file first to find them.