Mandelson Fleurival — Blog Developer. Designer. Builder. 2026-04-17T00:00:00.000Z https://mandelson.co/blog/ Mandelson Fleurival manny@mandelson.co Why I default to local models for bulk work 2026-04-17T00:00:00.000Z https://mandelson.co/blog/default-to-local-models/ Classify, summarize, extract. That's 80% of what agents actually do, and all of it runs fine on a laptop. Here's how I decide what goes local and what stays in the cloud. The bottom-up edit rule 2026-04-17T00:00:00.000Z https://mandelson.co/blog/the-bottom-up-edit-rule/ One edit always worked. Two edits usually worked. Five edits against a real file is where the flow fell apart quietly. The fix is a one-line rule in the system prompt. The 12,000-token message I didn't know I was sending 2026-04-16T00:00:00.000Z https://mandelson.co/blog/the-12000-token-message/ Jarvis's context was ballooning every time the reviewer cycle fired. The message looked tiny. Three-line critique, a few rules. I was actually shipping about 12,000 tokens per fire and had no idea. How I'd set up LM Studio today 2026-04-10T00:00:00.000Z https://mandelson.co/blog/how-id-set-up-lm-studio-today/ Local just crossed a line for me. The last Llama.cpp update bumped my token generation about 33% on identical hardware, and the gain holds across long agent sessions instead of collapsing after a few thousand tokens. This is the setup I'd copy today.