← barrosbuilds
Learning Loga public build log

Learn AI by building it. Every week.

I'm learning to build with AI in public — agents, MCP, RAG, evals, context management — one concept at a time. Each week I pick a tool, write the theory, build an exercise, and do it myself before it reaches you. Every exercise ships as a public repo.

You get the same thing I do: a weekly email with the concept and a hands-on exercise you can run. Plus what actually moved in AI over the past seven days, without the hype.

The exercises are yours to keep. They live on GitHub, they stay updated as the tools change, and they're designed to be applied to your own projects — not just followed.

I'm not an expert. That's the point. This is a build log, not a course from someone who already figured it out.

Free. Weekly. Unsubscribe whenever.
First issue ships October. No spam, no upsell.
Before you sign up

Where this starts from.

What I assume you have

  • →You've used a chat model — Claude, ChatGPT, something. I won't be explaining what a prompt is.
  • →You want to know whether the output is right, not just whether it looks right. “It looks fine” stops working fast, and you've noticed.
  • →You can read code when you need to. You won't write much of it — increasingly that's the whole point — but some bugs are ones you'd never have thought to test for.
  • →A terminal you're willing to open. About half the exercises are short scripts you run against the API; the rest happen inside Claude or Claude Code. Any editor works — VS Code is the easiest place to start. You'll also need an Anthropic API key, which is separate from a Claude subscription.
  • →Ideally, something of your own to point this at: a long chat you rely on, a repo you work in, a dataset you know well, a task you repeat every week. Every exercise ends by applying it to your own work. Nothing to hand yet? Each exercise comes with a sample to start on.

What you don't need

  • →To already know how to check the output. Tests, evals, comparisons against something you trust — setting those up is most of what this teaches.
  • →Any experience building with an LLM. If you've only ever used the chat window, you're exactly who this is written for — as long as you'll open a terminal.
  • →An ML background. No training, no fine-tuning, no maths. This is about building with these tools, not about how they work inside.
  • →To be a professional engineer. If you ship things with Claude and care whether they're right, that counts.

If none of that sounds like you yet, the exercises stay free and open on GitHub — come back when it does.

The curriculum

Thirty issues. Six tracks.

Every issue is one concept and one exercise you run yourself. The sequence runs from getting useful output out of a chat window to shipping an agent that runs on a schedule without you.

You start at issue 1 whenever you join — the sequence does not assume you read last week's.

  1. 01

    Working with the models

    Getting real output from the chat products.

    • 1What the model actually sees
    • 2Prompting as engineering
    • 3Structured outputs
    • 4Reusable setup
    • 5Context engineering
  2. 02

    Coding with agents

    Handing real work to a coding agent, and reviewing what comes back.

    • 6First real task
    • 7CLAUDE.md
    • 8Plan before code
    • 9Review loops
    • 10What not to delegate
    • 11Long tasks
  3. 03

    Knowing it worked

    Replacing “it looks good” with something you can measure.

    • 12Evals
    • 13Building a golden dataset
    • 14LLM-as-judge
  4. 04

    Extending the tools

    Connecting models to your own systems.

    • 15MCP
    • 16Using existing MCP servers
    • 17Building your first MCP server
    • 18Connecting a model to your own database
    • 19Skills
    • 20Why drop below the products
  5. 05

    Building your own

    Below the products, to the API.

    • 21The request loop
    • 22Tool use
    • 23Embeddings and vector search
    • 24Agents
    • 25RAG, end to end
    • 26Prompt injection and guardrails
  6. 06

    Agents that run without you

    Scheduled, persistent, and bounded — the ones you are not watching.

    • 27Retrieval quality and its limits
    • 28Memory that outlives the session
    • 29Delegation
    • 30Agents on a schedule

Every issue ships a public exercise repo. The concepts stay put; the exercises get updated as the tools change.