Practical thinking on AI context engineering, governance, and getting AI into production without regret. Written from the build side of the screen.
Most of the AI in a German insurer arrived rather than being built — through channels that never reach an inventory. What the 2025 supervisory reshuffle made visible, and where to start.
Read →A suggestion is not a constraint. My system prompt went from four flat headings to a gated multi-agent pipeline — eight versions, the same lesson.
Read →Why the questionnaire under-reports, and what to scan instead.
Read →The gains most teams chase in prompt wording actually live in what the model sees. A practical look at engineering context.
Read →Treating context as a scarce, designed resource — what to include, what to leave out, and why it matters for reliability.
Read →