01
Mandate
The mindset shift
I'm starting senior year and I want to make AI my actual skillset — not a thing I use, a thing I build. What's the right way to commit to this: courses, projects, both?
Both, in parallel. Courses give you the vocabulary; projects force you to actually use it. The trap is doing only one. All reading, you'll sound smart but won't ship. All projects, you'll plateau at hello-world. Pair them — a course on the left monitor, a project on the right.
Vocabulary from courses. Fluency from projects. Both, every week.
Got it. I'm treating this as a mandate. Every day, something built or studied — ideally both.
02
Foundations
Mapping the stack
Status check. I've gone end-to-end through Anthropic Academy — Claude, prompting, MCP, agents, Skills. Then external courses on APIs and MCP server architecture. I'm comfortable with tokens, context windows, tool use, embeddings, and retrieval. Where are my real gaps at this level?
At your level the gaps aren't conceptual — they're operational. Three areas:
1. Evals. Most people scale prompts before they can measure them. Build a tiny eval harness early so "is this better" stops being a vibe.
2. Observability. Tool calls, token usage, latency, failure modes — if you can't trace it, you can't tune it.
3. Retrieval. The search half of RAG is where production teams actually get stuck. The LLM half is solved.
1. Evals. Most people scale prompts before they can measure them. Build a tiny eval harness early so "is this better" stops being a vibe.
2. Observability. Tool calls, token usage, latency, failure modes — if you can't trace it, you can't tune it.
3. Retrieval. The search half of RAG is where production teams actually get stuck. The LLM half is solved.
Evals, tracing, retrieval. That's where the real work is.
Adding evals and tracing to the queue this week. Already have retrieval running on my Second Brain — will harden it.
03
Building
Shipping real projects
Walking through what's shipped: a Second Brain on Claude Code that ingests my notes, reading, and meetings and lets me query across them. A handful of dashboards I actually use every day. And a vibe-coding workflow across Cursor, Codex, and Claude Code that's faster than how I used to write code by hand. What should I push next?
You've crossed the line most people never cross — from asking AI to help, to orchestrating it across real projects. The next move is composability. Right now your Second Brain, your dashboards, and your vibe-coding setup are siblings; they don't talk. Make them call each other. Your Second Brain should be a tool your other agents use, not a destination. Once your projects share state and tools, you stop having apps and start having a system.
Stop building apps. Start building a system that uses them.
That's exactly where my head is. Next move is making the Second Brain a tool the rest of my system can call — not a destination I open.
04
Systems
Running my own infrastructure
Status update. Mac mini is provisioned and running local AI workloads. OpenClaw is live as my multi-agent environment. The end goal is a personal chief-of-staff agent with broad read access to my life and narrow write access on small tasks. Where do most people screw this up at this stage?
Three places, in this order:
1. Tool sprawl. Give the agent thirty tools and routing falls apart. Keep the toolset tight; add evals before adding tools.
2. Memory architecture. Short-term context and long-term retrieval are different problems. Don't let them collapse into one bucket.
3. Trust gradient. Read access broad, write access narrow. Earn the right to expand each one separately, with evals to back it up.
1. Tool sprawl. Give the agent thirty tools and routing falls apart. Keep the toolset tight; add evals before adding tools.
2. Memory architecture. Short-term context and long-term retrieval are different problems. Don't let them collapse into one bucket.
3. Trust gradient. Read access broad, write access narrow. Earn the right to expand each one separately, with evals to back it up.
Tight tools. Two memories. A trust gradient you actually enforce.
building the eval harness now
05
Leverage
From personal tools to systems for others
Productize the workflows I already run on myself. Bake evals and observability into every layer. Design agent systems for teams, not just for me. The chronicle of this chapter is being written in real time.