The DevOps Learning Command Center dashboard existed. The problem was keeping it alive.

A learning system that requires manual maintenance eventually stops being maintained. I wanted Claude Code — the terminal-based AI agent from Anthropic — to handle the vault upkeep so I could focus on the actual learning.

What Claude Code Does

Claude Code runs in your terminal and has direct access to your filesystem. It reads files, writes files, runs commands. Combined with a CLAUDE.md context file that explains who you are and what the vault is for, it becomes something like a personal assistant that actually knows your project.

The key concept is skills — markdown files stored in .claude/commands/ that define exactly what Claude should do when you type a slash command. Think of them as precise instructions for specific recurring tasks.

The First Skills

The first skill I built was /save — a single command that commits all vault changes and pushes to GitHub. Dead simple, but it removed the friction of remembering to back up.

Then /wiki — a knowledge base workflow based on Andrej Karpathy’s note-taking method. When I learn something worth keeping permanently, /wiki helps me extract it into a structured, indexed note that’s findable later.

Then /morning-briefing in its first form — load today’s context, show what’s planned. This one went through several iterations before landing somewhere useful. More on that in a later post.

What Didn’t Work

The first version of Claude Code vault management had a fatal flaw: I asked it to do too much. Reschedule tasks. Sync calendars. Update progress counters. Generate review schedules.

Every operation read multiple files, made multiple tool calls, burned through tokens at a rate that added up fast. One /reschedule run could cost more than a day’s worth of learning sessions if I wasn’t careful.

The lesson was to keep skills narrow. Each skill does one thing. No searching, no exploration — direct file reads by exact path, one calendar call maximum. The vault is structured enough that Claude never needs to hunt for anything.

The CLAUDE.md File

The most important file in the whole setup is .claude/CLAUDE.md. It’s loaded automatically at the start of every Claude Code session and contains everything CC needs to know: who I am, what the vault structure looks like, what each skill does, and — critically — what it should never touch.

That last part matters. The calendar policy section now reads: Claude Code does NOT manage calendar scheduling. User owns the schedule. That boundary took a few expensive lessons to establish.

Next: choosing a task manager, and why three different tools each taught me something different.