50 sessions on gradio with a coding agent: the agent failed at least once in 33, and 14 finished without a human stepping in.
Takeaway: Next time I'd lead with the Space's hard constraints (1GB repo limit, free-tier build memory, no CUDA) before letting the agent draft a plan, since two of our failures were pure packaging/build issues we could have anticipated upfront.
Takeaway: Next time I'd just ask for the mode buttons themselves to change color from the start, since that's the solution I actually wanted — the badge approach was wasted work.
Takeaway: I'd start by asking for a real browser-based inspection (Playwright/DOM dump) from the very first CSS complaint instead of letting the agent guess at selectors screenshot by screenshot - that back-and-forth ate most of the session.
Takeaway: Next time I'd ask upfront for a deploy script instead of letting the agent hand-run git/HF CLI commands each time — that's what actually stabilized the workflow.
Takeaway: Next time I'd start by telling the agent explicitly not to commit anything without asking, since it committed twice before I stepped in to stop it.
Takeaway: I'd start by pointing the agent straight at the specific prize categories I care about instead of asking for a general read of all the docs, since most of the session was just it catching up on context I already knew.
Takeaway: Next time I'd start by asking for a ground-truth labeling workflow and benchmark harness before building any alignment logic, since that's what actually let us tell good methods from bad ones.
Takeaway: Next time I'd specify upfront that word clips must not bleed into adjacent words, since that was the real bug hiding behind 'alignment isn't better' — the fix was clamping word end to the next word's start plus much smaller padding for aligned clips versus the proportional fallback.
Takeaway: Next time I'd start the first prompt by stating explicitly that this is a kids' reading app needing realistic per-word audio with strict latency and reliability constraints, so we don't cycle through browser-TTS-vs-VoxCPM-vs-Modal three separate times.
Takeaway: Next time I'd open with the fact that TTS needs to work instantly for a kid clicking words, not just eventually — that would have surfaced the Modal cold-start problem before I burned time on Gradio 6 launch-arg compatibility.
Takeaway: Next time I'd ask for the environment sanitization (Conda vs .venv Torch conflicts) to be handled up front, since that same libtorch_python.so issue bit me twice, once during GGUF conversion and again during local_inference.py testing.
Takeaway: Next time I'd give the exact test fixture (silent WAV) up front so the agent designs the endpoint to handle it from the start instead of hitting a gated-model load in production.
Takeaway: Next time I'd specify upfront that the app runs on Gradio/HF and dark mode might be triggered by a class rather than just OS preference, since that's what forced a second round of patching.
Takeaway: Next time I'd ask upfront for the prompt's exact output-format constraints, not just the risk-level instructions, since the rigid 4-week structure was quietly forcing intervention-heavy plans regardless of risk.
Takeaway: Next time I'd ask for the risk-group breakdown (Low/Moderate/High/Critical) from the very first scoring pass instead of asking for it turn by turn, since that grouping is what actually revealed the real problem was the fixed 4-week output format, not retrieval or root-cause analysis.
Takeaway: Next time I'd specify upfront that action plans must be pasted in manually and that nothing should call the model automatically — I went through three different generation strategies (UI export removal, background auto-generation via Qwen, then placeholder-only extraction) before landing on what I actually wanted.
Takeaway: I'd start by telling the agent upfront that the local venv is missing deployment-only packages (torch, transformers, spaces) so it doesn't waste time discovering that itself.
Takeaway: Next time I'd open with both asks together — what goes into the embedding and why there's a 100-project cap — since the agent had to dig through the same files twice.