I Used Claude Code for 8 Months — Here's What Actually Stuck

What you’ll build

A three-agent Claude Code setup — one agent for your personal brand, one for your business, one for your personal life — where each agent has its own memory vault and a library of reusable skills you invoke instead of doing the work by hand. This lesson gives you the exact skills I demoed, free to download and drop into your own setup.

What you’ll need

• Claude Code installed
• An Obsidian vault per agent (memory / context store)
• A Notion account (for the idea + task skills)
• A Google account with Fathom connected (for the meeting skills)
• The free skill repo — link in Resources

Step 1: Split your work into three agents

Don’t run one do-everything agent. Give each domain its own agent with its own CLAUDE.md and its own vault:

• Axiom — personal brand: diagrams, Notion ideas, content editing + scheduling
• Atlas — business operations: call prep, meeting notes, contracts, cold email, books
• Jarvis — personal life: journaling, open loops, finances, morning brief

Four universal skills live in all three (/wrap, /superpowers, /skill-builder, /llm-council). Everything else is agent-specific.

Step 2: Give every agent a vault

Each agent is connected to its own Obsidian vault — a Wikipedia of everything you’ve done in that domain. This is the real unlock. A session only remembers what’s in CLAUDE.md, and answers degrade past ~200K tokens. A vault feeds the agent deep context on demand, so builds and answers come out 10x better.

Keep the vaults separate: brand knowledge (hooks, scripts) in Axiom’s, client/pricing/meeting knowledge in Atlas’s, journal + personal in Jarvis’s.

Step 3: Turn rambles into structured Notion entries

Install /notion-brain on your brand agent. When a video idea hits you, ramble at it for 5–10 minutes and it structures the whole thing into your Notion idea database — difficulty, form, status, tags, and a page outline you can build diagrams, thumbnails, and scripts from.

Axiom, ingest this video idea into the Notion database: <ramble>

Step 4: Generate architecture diagrams in seconds

/excalidraw-diagram builds video-grade diagrams with real logos and self-renders them to PNG so you can QA before using them. Point it at a concept and it draws it.

Excalidraw, draw a diagram of my content repurposing pipeline

Step 5: Automate meeting notes and prep

Connect Fathom to your work email so it joins every Google Meet and records a transcript. Then:

• /meeting-notes pulls the transcript from Fathom and outputs key takeaways, decisions, and action items — drops the doc in Drive and pushes tasks to your team hub. No note-taking during the call.
• /meeting-prep pulls every open task from the team hub into a talking-points doc before your weekly internal call, and confirms with you before it builds.

Together they make a loop: notes create tasks → next week’s prep checks those same tasks got done.

Step 6: Journal + track open loops with Jarvis

At the end of the day, message Jarvis (I do it from Telegram) with what you got done and what’s still open:

Just wrapped recording the skills video. Took way longer to set up the demos than I figured. Open loop: still need to finish editing.

It appends to your journal vault and tracks the loop status. Open loops = anything nagging at you that isn’t finished. Close them to free up mental energy.

Step 7: Add the universal skills

These four run under everything — install them first:

• /wrap — end of a build session, this saves the lessons and memory so the agent actually gets better instead of starting from scratch every time. This is how I fixed meeting-prep: reiterate, then /wrap.
• /superpowers — brainstorm → plan → build → verify rails. Invoke it for anything complex (a portal, an app) so it plans before it builds. Ships from a public repo (link in Resources).
• /skill-builder — the skill that builds skills, in the structure Claude invokes best. Use it for every new skill.
• /llm-council — five frontier models (via OpenRouter) argue a big decision from different angles, then Claude chairs the final verdict. Kills the yes-man problem.

Step 8: Stress-test plans with /grilling

When plan mode only asks 2–4 questions and you want more rigor, /grilling hammers you with 10–20 questions to nail down exactly what you’re building before a single line gets written.

Step 9: Run all three at once

The point isn’t any one skill — it’s running the agents in parallel and swapping between them while they work. A real morning: Jarvis runs your morning brief, Axiom generates a thumbnail, a finance skill reconciles your cards, meeting-notes runs during your call — all at the same time. You’re invoking skills, not doing manual work.

Resources

• GitHub repo (same skills, public): https://github.com/austinpeechatt/claude-code-starter-skills
• Superpowers (universal skill, public repo): https://github.com/obra/superpowers
• Wispr Flow — voice-to-text I ramble every skill into (1 free month): https://wisprflow.ai/r?AUSTIN16480