How Olddogs.ai Built Its Site with Agents Without the Slop

Summary

Olddogs.ai describes how its homepage was built entirely through prompts to AI agents, while a set of lightweight tools helped preserve a distinctive design. The team used a design-build loop: generate five to ten genuinely different visual directions with generative image models, refine selected ideas, have an agent implement them in code, and compare screenshots with the chosen mockups. The process was repeated across eight site sections, producing 2,460 generated images and multiple design/build rounds. The team also created an in-app design jig, a development-only control panel for changing parameters such as density, type scale, layout, and animation behavior, then exporting the selected settings back to code. A separate standalone jig reviewed and refined more than 50 pairs of calm and activated dog images, with one-click notes exported to Codex. For performance, the team asked its agent Astra to rewrite a canvas-based cell simulation as a GPU shader, raising the measured frame rate from 14 to 117 fps during fast mouse drags in a 1,440 by 900 test on a 120 Hz display. Astra also added five shader-based dog animations, increasing the page by about 77 KB. The same tools and workflows were used to build the blog post and its interactive demonstrations, and the team has shared a downloadable skill and prompts for reuse.