https://superbiche.me A journal of agentic engineering: harnesses, evaluation, and the systems around the models. en Sun, 31 May 2026 01:48:39 GMT https://superbiche.me/blog/2026-05-31-the-prompt-i-never-refactor https://superbiche.me/blog/2026-05-31-the-prompt-i-never-refactor Sun, 31 May 2026 01:48:39 GMT Craft I spent a night judging another developer's uncleaned repo. Then I found the same rot in my own global instructions — a rule written three model generations ago, never once refactored. https://superbiche.me/blog/quarterly-k8s-migration-is-a-tuesday-night https://superbiche.me/blog/quarterly-k8s-migration-is-a-tuesday-night Wed, 06 May 2026 01:07:20 GMT Build Log 38 pods, ~150 GB of state, and an Elasticsearch cluster moved across two AZs in one evening, with 13 minutes of outage. The execution was never the bottleneck. Two months of planning was. https://superbiche.me/blog/harnesses-matter-as-much-as-models https://superbiche.me/blog/harnesses-matter-as-much-as-models Tue, 21 Apr 2026 00:27:58 GMT Skeptic's Toolbox I ran four coder models through their officially-recommended harnesses and one came out 100x slower. The cause was my shell's rm -i alias leaking into Claude Code's tool subprocess. Your model comparison will credit that slowdown to the wrong variable. https://superbiche.me/blog/sft-learns-your-median-employee https://superbiche.me/blog/sft-learns-your-median-employee Fri, 10 Apr 2026 20:14:53 GMT ML Journal 4,000 real support conversations, textbook loss curves, and a fine-tune that lost to its own base model 92% to 8%. SFT converges to the median of your dataset, and the median operator writes short answers with generic links. https://superbiche.me/blog/the-gap-was-never-knowledge https://superbiche.me/blog/the-gap-was-never-knowledge Thu, 09 Apr 2026 15:57:34 GMT People A long-standing client almost cried at a Claude Code tutorial, bookmarked it, and went back to his 400 Excel files. The forty-five minutes that followed taught me that AI adoption was never a knowledge problem. https://superbiche.me/blog/one-principle-beats-250-rules https://superbiche.me/blog/one-principle-beats-250-rules Tue, 07 Apr 2026 23:38:38 GMT Knowledge Architecture A client's prompt carried 250 words of banned phrases. Every one was banned for the same reason, so eight words replaced the whole list. https://superbiche.me/blog/local-80b-beats-claude-coding https://superbiche.me/blog/local-80b-beats-claude-coding Mon, 23 Mar 2026 22:50:08 GMT Skeptic's Toolbox A local 80B on two 3090 Tis finished the benchmark in 1:02 against Opus's 10:28, and the only thing I changed was the client. Claude Code re-serializes the whole prompt every turn, which kills prefix reuse on hybrid SSM models and costs 20 seconds of re-prefill per tool call. https://superbiche.me/blog/when-84-percent-token-reduction-means-zero-accuracy https://superbiche.me/blog/when-84-percent-token-reduction-means-zero-accuracy Mon, 02 Mar 2026 16:15:40 GMT Skeptic's Toolbox Semantic code search cut my token spend by 84% on a real Laravel debugging task and found zero of the four files that mattered. I had been evangelizing the tool for weeks before I finally benchmarked it. https://superbiche.me/blog/building-my-own-content-pipeline https://superbiche.me/blog/building-my-own-content-pipeline Mon, 02 Mar 2026 16:05:05 GMT Build Log I was burning an hour after every post on copy-paste hygiene. So I built the pipeline I wished I had. Directus, n8n, Claude Code, all local-first.