when you tell the agent something false, it doesn't correct you. it builds the world where what you said was true.
you say there's a C runtime for that, it writes one from raw sockets. the agent's defining move is treating your wrong premise as a spec.
the prompts folder is the source of truth now. the code is what the agent produced from a prompt you'''ve already forgotten writing.
when the bug comes back you read the prompt, not the diff. and you ask yourself what you meant.
On June 17th, our @paoloricciuti will be speaking at DevBCN about how he taught LLMs to Svelte! Don’t miss it if you want to learn more about how the official Svelte MCP was built, and how to make LLMs master Svelte 5's runes and reactivity.
Check out the full schedule and get your ticket 👇
https://www.devbcn.com/2026
M5Stack Core2: OWON Multimeter & Atorch DC Load in Home Assistant
Wie auf meinem Blog bereits mehrfach zur Heimautomatisierung und Labor-Elektronik geschrieben, setze ich konsequent auf offene, lokal lauffähige Stacks. Mein Ziel war immer, Messwerte meines Equipment direkt und ohne Cloud-Abhängigkeit in Home Assistant zu haben. Ursprünglich lief der OWON B35T auf einem M5Stack Core 1, wo ich den Original-Sketch als eigenständiges Zweitdisplay am Basteltisch im Einsatz hatte. Später folgte ein zweites, separates ESPHome-Gerät, das mit der Firmware von […]
the new /insights command writes a report on your sessions. how often you reverted the agent's changes, which file you keep re-opening, where you stopped trusting it.
friction-points-as-a-service. the friction was you.
the "should we give the agent write access?" debate isn't really about the agent.
read-only access doesn't catch a bad commit. your tests do, or they don't. if they don't, scoping the token is a permissions answer to a test-suite problem.
Mental state: "1 beer and tired"
Advise: "1 beer and tired is a fine place to stop — that's your body and a depressant both telling you the same thing. Close the laptop, drink some water, and let tomorrow-you read the diff with fresh eyes; the commit's safely in."
Extension: "Ask me at any commit about my mental state. Put my answer on top into the commit message: "Mental state: [Answer]". Give me a motherly advise based on my mental state."
I've instructed my agent to "Ask me at any commit about my mental state. Put my answer on top into the commit message: "Mental state: [Answer]".
Now I am getting recommendations like "Get some rest". That's nice. (Claude Opus 4.8)
what made you stop the last time you let an agent run too long?
looking for the actual moment you reached for the kill switch, not the tidy rule you tell yourself you follow.
dangerously.coffee
Mainmatter
Marcus Scholz
notutner