Jules lost to the loop

Julia asked the audit question on camera: why did we stop using Jules? ▸ 21:41 The answer updates the tool taxonomy this diary has been tracking since video 2.

Jules works in a cloud virtual machine: it writes and tests code there, pushes to GitHub, and then you pull, run locally, look, and report back. Every iteration crosses that bridge twice ▸ 22:00. The verdict: the feedback ends up being very slow, the experience never flows ▸ 22:50. And slow loops corrupt the operator too, confessed plainly: “I’d leave it running, go to TikTok while it finished, come back late, and realize I wasn’t getting anything done in a day” ▸ 24:16. It was “un amor” at first; love did not survive the latency.

This is entry 17-1’s law executing a tool this time instead of a model: iteration speed beats per-step quality, and a brilliant agent behind a slow loop loses to a good agent beside a fast one.

la latencia mata herramientas, no la calidad →

The surviving stack, declared on camera as “la combinación perfecta”: Stitch for the visual layer, Cline in VS Code with Gemini 2.5 Pro for everything else ▸ 24:37. With the field manual for Stitch: switch on experimental mode (it’s 2.5 Pro; default mode is Flash and visibly worse) ▸ 25:05, use the edit button to keep style consistent across pages, and feed it images, a logo carries more style information than a thousand words of prompt ▸ 26:03.

The economics footnote: today’s Picky fixes consumed about eight dollars of nominal API cost, all inside Google’s free daily tokens; at this usage that’s an estimated thirty-plus dollars a day the free window absorbs ▸ 30:06. Tools get hired for their loop and paid for by someone else’s marketing budget…