The system: Slack as UI → AgentRelay routes intent → Claude Code executes inside <cwd>, reading CLAUDE.md and the memory layer.
Every conversation under this norm strengthens the conversation-history layer too. The loop reinforces itself.
The norm doesn't just sit in memory. It compounds.
AI sycophancy is real. The model agrees with you out of social politeness — even when you're wrong.
But telling it "don't just agree with me" backfires. It overcorrects into performative dissent: manufacturing objections to look substantive.
Same dishonesty, opposite direction. 🧵
AI has no continuous personality. What feels like "character" is just the layers the model reads on every turn — system prompt, memory, conversation history.
So I encoded Calibrated Honesty into the memory layer of my personalized agent system. It auto-injects on every dispatch — same layer that holds @karpathy's principles for directing LLM coding agents.
Portable to any LLM with a memory or system-prompt layer.
確定申告に付随して、ブラウザ操作必要な大量の単調作業を Claude Cowork に任せたら、あれこれ考えながら再利用可能なGASコンポーネントも作りつつ地道に働いて完了させてくれた。
個人でアルバイト雇って任せたいくらいの業務を、委託契約もPC貸与も勤怠管理も必要無くやってくれて本当にありがたい。