AI Learning Fleet

A small team of AI agents that teaches its owner to think like an AI product manager — and keeps itself running, every day, with no babysitting.

Most "I'm learning AI" projects are a folder of bookmarks nobody reopens. This is a system: a handful of cooperating agents that work out what you should learn, teach it at your level, make sure it sticks, and keep you current — automatically. It was built by a product manager, on Claude Code, to go from "I can follow AI headlines" to "I can reason about AI products" — and to be honest, working proof of that journey.

---

The problem it solves

Learning AI from a non-technical start fails quietly in two ways:

• You don't know what to learn, or in what order. The field is enormous, bookmarks pile up, and nothing connects.
• What you read doesn't stick. You understand an article on Tuesday and can't explain it by Friday.

The Fleet is designed to close both — deliberately, and with as little willpower as possible.

The fleet

Each agent does one job; they all share one memory (the Knowledge Ledger — a running model of what you know).

Agent	What it does, in plain English
Cartographer	Maps the whole landscape of AI concepts a PM should know — in sensible order — and shows you your blind spots.
Decoder	Paste any article or link; it explains it at your level — the prerequisites you're missing, the jargon decoded, why it matters for a PM, and a few questions to check you got it.
Scout	Every morning, pulls your chosen sources, picks the 1–2 most relevant pieces, and digests them into a two-minute brief.
Drill	Quizzes you on what you've learned using spaced repetition, so it actually sticks — and only marks a concept "known" once you can explain it unaided.
Conductor	Runs the whole thing on a daily schedule, logs every run, and keeps it all backed up. No babysitting.
Showcase (planned)	Turns your progress into shareable write-ups and a skills-vs-role-requirements view.

flowchart LR
  C[Cartographer<br/>what to learn next] --> L[Decoder &amp; Scout<br/>learn it]
  L --> K[(Knowledge Ledger<br/>what you know)]
  K --> D[Drill<br/>make it stick]
  D --> K
  K --> C
  CO([Conductor]) -. runs it daily .-> L

Loading

The point isn't any single agent — it's that they compound. The more you use it, the more the system learns about you, and the better every agent gets at meeting you where you are.

Why it's different

• A system, not a reading list. It decides what's next, teaches it, and verifies you retained it — closing the loop instead of just adding to a pile.
• It learns about you. Every session updates a shared memory of what you understand, so explanations stop repeating what you already know and focus on your real gaps.
• It runs itself. One daily job, ~$0 to operate, no servers — it produces your brief, schedules your reviews, and backs itself up unattended.

What it feels like to use

• 7:30 a.m. — a brief lands: two relevant reads, the next concept to learn, and how many reviews are due.
• Five spare minutes — learn the next concept, quiz me, or decode this article.
• Over weeks — your concept map fills in from the foundations upward, and ideas move from seen → understood → can explain.

Under the hood (for the technically curious)

Deliberately simple, cheap, and robust:

• Runs locally on Claude Code with a personal subscription — ~$0 marginal cost, no servers, no third-party platforms.
• A shared, human-readable memory (plain JSON/Markdown, version-controlled) tracks every concept's status, your comprehension level, and its review schedule.
• Combines content retrieval (RSS + article extraction), spaced repetition (Leitner scheduling), and safe automation (scheduled runs, file-locked shared state, graceful per-agent failure handling, nightly auto-commit).
• 42 automated tests, self-documenting run logs, and a written requirements + plan history in docs/.

Concepts exercised: AI product management · agentic workflows · LLMs · RAG · embeddings · semantic search · evaluation · prompt design · retrieval · spaced repetition · automation · Python.

Status

Live and running daily: Cartographer, Decoder, Scout, Drill, Conductor. Showcase is planned. Each agent was shipped and tested before the next was started.

Run it yourself

# from the project root, in Claude Code
py -m venv .venv && .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python scripts\doctor.py        # health check

Then /map to see the concept map, /next to learn one, /decode <url> to digest an article, /drill to review. See docs/ for the full requirements and plan.

---

What this project demonstrates

Kept to the end on purpose — the work should speak first.

It started as a way to learn AI. Building it ended up exercising the core of the product role:

• Requirement clarity — it began as a written brief and plan (problem framing, scope boundaries, success criteria, explicit non-goals) before any code; that paper trail is in docs/.
• Product thinking — sequencing by value and dependency (ship what helps on day one, defer the rest), holding scope, and building the smallest version that delivers real value.
• AI fluency — embeddings, retrieval/RAG, evaluation, agentic orchestration, prompt design, and the practical build-vs-buy tradeoffs of putting models into a product.
• Systems judgment — a self-correcting feedback loop, graceful failure handling, cost control, and "deploy and maintain" rather than "demo once."
• Communication — explaining technical ideas at the reader's level, which is, after all, the entire product.

Built and run by Likitha Madala.