Daclaude

A personal AI agent system built on ~30 Claude Code skills that orchestrates daily life: email triage, daily recaps that synthesize activity across tools into a structured journal entry, weekly reviews that process a notes inbox and recommend what to write next, content digests from newsletters, a quest system for task management, finance coaching with live data from personal finance tools, and more. No custom code — it’s prompt engineering, a notes app as persistent state, and tool integrations for external connections.

The interesting part isn’t the automation list. It’s that the system improves itself through use. Each skill is a SKILL.md file — the prompt IS the codebase. When a skill doesn’t work well, you edit the prompt, and the system gets better. The weekly recap skill was upgraded mid-session during a review to include content recommendations for my personal website. The handoff skill captures learnings at the end of each session. The adversarial-verifier catches mistakes. Shared context flows through auto-memory (cross-session learning) and episodic memory (conversation history search), so insights compound.

%% TODO: Add specifics on what works well — email triage time savings, daily recap journal consistency, quest visibility. Need before/after comparison of workflow pre-Daclaude. %%

%% TODO: Add honest section on what doesn’t work — auth breaks constantly (Gmail OAuth, API tokens), skills can’t coordinate mid-session, memory is append-only not curated, system is brittle to vault structure changes. %%

The meta-lesson: building personal tools with AI agents is the best way to develop intuition for building AI products. Every limitation I hit with Daclaude — memory management, tool coordination, auth brittleness, the gap between “works in demo” and “works daily” — teaches me something about agent architecture that applies directly to the products I build at work.