Notes from the workshop.
Engineering posts on what we're actually building — prompt science, brand DNA, render pipelines, and the work behind a content calendar that runs itself. Honest about the parts that work, the parts that don't, and the numbers we can measure.
Latest.
Auditing an agent's memory — six silent failures behind "she actually learned something."
An agent's memory layer can look alive — extract, compress, surface — and still be silently broken end-to-end. The audit Desklight ran on Allie's persistent memory, the six failure modes hiding behind successful-looking log lines, and the moment the layer started telling us something we didn't already know.
Sweet spots, not ceilings — how Desklight adapts prompts for every model.
Long prompts don't truncate, they drift. The brand DNA is the water; each model is a different vessel. The per-model prompt adapter reshapes brand DNA to fit the active model's sweet spot before the call ships — same water, every vessel, every render. With citations from Lost in the Middle (TACL 2024) and DetailMaster (2025) on why effective context is tighter than published context.
Prompt science is the product — Desklight's translation compressor.
How we treat prompts like code: the translation compressor that turns one calendar entry into a brand-locked render, the per-model dossiers we maintain across the current image and video frontier, and the validators-as-code pattern that keeps Allie honest.