Dev3Pack AI-Engineering Bootcamp

Three weeks. 15 sessions. One source-grounded research assistant you can test, cite, and defend.

14 September – 2 October 2026, Monday to Friday, two hours a day. Optional showcase on Saturday 3 October.

Contents

• Start here
• Get started with Claude Code
• Find your way
• For coding assistants
• What you will learn
• Units, sessions, extras
• Repository map
• Course map
• What arrives, and when
• Week 0
• Weeks 1-3
• Final assignment and certificate
• Bonus and extras
• Commands
• Safety

Start here

git clone https://github.com/Gecko-Academy/dev3pack-cohort-2026-09.git
cd dev3pack-cohort-2026-09
uv run bootcamp start          # installs everything, then says whether it worked

start is safe to run as many times as you like, and it ends by telling you the one command to type next. On macOS and Linux make setup is the same thing.

Windows: run these in Git Bash or WSL2, not PowerShell. The repository is public — there is nothing to request and nothing to wait for.

If you have SSH keys set up, git clone [email protected]:Gecko-Academy/dev3pack-cohort-2026-09.git works too. HTTPS is above because it needs nothing configured first.

Open 00-START-HERE.ipynb. It lists every notebook in order and ticks what you have finished.

New to the repo? SETUP.md is the half-hour path to a green doctor. No API key is required: every scored notebook runs offline against FakeLLM. This repository is read-only for you — hand in to the submissions repository instead (see Handing in); week 0 is here now, later weeks land on 14 / 21 / 28 September — git pull at the start of each week.

Get started with Claude Code

Claude Code is an assistant that runs in your terminal. It reads the files in the folder you start it in. This repository carries seven skills and three subagents written for this course. They are already in your clone. Nothing to install and nothing to switch on. A git pull keeps them current.

Optional, and never counted. Every scored exercise runs offline, with no key and no assistant. Sessions 1 and 10 are the two that want one open.

Install it, then start it in the course folder

npm install -g @anthropic-ai/claude-code   # once; needs Node.js
cd dev3pack-cohort-2026-09
claude

The folder matters. It is everything the assistant can see. Start Claude Code anywhere else and none of this loads.

To leave, type /exit. To start a fresh conversation, /clear. Long conversations drift.

What loads by itself

Claude Code reads CLAUDE.md on startup. The first line of that file is @AGENTS.md, which is an import and not a link, so AGENTS.md loads with it. That is the policy that stops an assistant writing your exercise for you.

Check it worked: type /context and look for CLAUDE.md in what it lists. If it is not there, you are in the wrong folder.

How to use a skill

A skill is a short set of instructions for one job. Describe your job in plain words and Claude picks the matching skill by itself. The first line of each skill says when to use it, and that line is how it chooses.

my check for session 4 is failing and I do not understand the output

You can also name one outright, which is useful when it picks the wrong one:

use the check-my-exercise skill on ch04

How to call a subagent

A subagent is a separate helper with its own, smaller set of tools. Ask for it by name:

ask the session-tutor to teach me session 6

The tutor can only read and search files. It has no tool that writes. So it cannot change your notebook, not even by accident, which is the point of it.

One worked example

Type this:

study session 6 with me

Claude loads the study-a-session skill, reads every page in units/en/unit2/session-06-retrieval-baseline/, and explains retrieval one idea at a time. Every claim names the page it came from, like units/en/unit2/session-06-retrieval-baseline/concepts-2.mdx, so you can check it yourself. Then it stops and asks you two or three questions, and waits for your answers.

When you are ready, it points you at that session’s notebook.ipynb and stops there. You write the exercise.

What ships

Skills (.claude/skills/):

Subagents (.claude/agents/):

What not to ask it for

The answer to a graded cell. Every skill and subagent here refuses it, and so does AGENTS.md. They explain the idea, name the page that teaches it, and leave the writing to you.

That is not a rule for its own sake. At demo day you explain your own work. An exercise your assistant wrote is one you cannot defend, and the marks were never the point.

Also worth knowing: never paste a key, a token or a .env file into any chat, and never accept a diff you have not read.

This kit and the plugin are different things

builder-kit/ is a Claude Code plugin. You install it, it works in any repository, and its commands are namespaced.

This kit is the opposite. It is files in the repository you already cloned, it works only here, and it is loaded the moment you start Claude Code in this folder. Use both if you like. They do not collide.

Find your way

Looking for today’s exercise? It is the notebook.ipynb inside the session’s folder. Every link below opens the file itself, not a folder to search.

Inside a notebook, everything that runs as shipped is already done for you. Each session has exactly one cell that does not — its heading says “Challenge”, and it tells you what it is worth, what done looks like, and gives you a tip.

Handing in

git pull                                              # always first: fixes reach you this way
uv run bootcamp check ch04                            # the scorecard, before you submit
uv run bootcamp submit ch04 --github  --push     # opens the pull request for you

Submissions go to Gecko-Academy/dev3pack-submissions, never to this repository. In the pull request’s Files changed tab, every file should start with submissions//. It merges on its own once its check is green.

For coding assistants

Read AGENTS.md first. CLAUDE.md and .cursor/rules/bootcamp.mdc point at it — do not fork the policy.

To answer a question about the course, search it. The course MCP server at https://mcp.geckovision.tech/course/mcp serves every published page with no key. Cite the page_id of each passage you use. An empty result means the course does not cover it: say so instead of guessing.

Then read in this order:

Inspect → plan → implement → test → review. No new dependency unless a lesson asks for it.

Optional, local only — never required, never committed:

uv tool install graphifyy
graphify install

Then /graphify . in your assistant. Leave graphify-out/ untracked.

What you will learn

• Harness — assistant instructions, traces, evals, and CI around the model
• Contracts — one adapter, typed outputs, bounded read-only tools
• Loops — budgets, stopping conditions, refusal as a first-class exit
• Grounding — retrieve from a versioned corpus, cite sources, refuse the unsupported
• Surfaces — MCP, deployment, and defending a failure from your own traces

Three threads recur: loop, graph, and harness engineering. The full day-by-day plan is in docs/curriculum.md.

Units, sessions, extras

Two sessions are assistant-driven (1 and 10) and show as in Jupyter in bootcamp progress.

Repository map

Where everything lives, and what a student receives.

7 paths are withheld by name, whatever the week: .github, docs/instructor, docs/plans, docs/specs, evals, scripts/instructor_pack.py, tests. Each is on that list for a stated reason — tests/ alone holds the solved value of every exercise. The publisher audits the tree it built rather than trusting the copy, and refuses if one of them appears.

Sent in part: docs/, modules/. Guides, the curriculum and the generated index ship; the instructor material, the specs and the plans do not.

Course map

What arrives, and when

This repository grows during the course. Week 0 is here now; each Monday one more week lands and you git pull.

Nothing is hidden behind a permission. A week that has not opened is not in the repository yet, so git pull on the Monday is the whole ritual. Your table of contents lists what you actually have, and names what is still to come in a comment at the end — this table is the schedule, and you have it from day one.

A correction is withdrawn as well as added: the publisher removes a file the current week no longer contains, so a fix actually reaches somebody who already pulled.

Solutions: week 0 and sessions 1–3 carry a solutions/ notebook. From session 4 on, they do not. Each session ships with everything running except one cell, and that cell is yours to solve — a solution would only be that cell, answered. When you are stuck, hint(reveal=True) is the worked answer (it costs marks), and coach(...) answers from the course pages for free.

Week 0

Self-paced. Handed in as a record, never marked. Short on time? Do the four 45-minute units. Courses A–C stay here for the whole course and nothing in the fifteen sessions waits for them.

Fast lane

Course A — Software engineering foundations

Course B — MCP

Course C — Data structures for agents

Weeks 1-3

Each live session is one directory under units/en/unit1/, unit2/ or unit3/. Exercise opens the notebook you work in; Demo is the one run in class. A session that has not happened yet may not be in your copy — git pull.

Session 13 uses an instructor-hosted Gecko MCP surface; the URL is handed out in class.

Final assignment and certificate

The final assignment is a source-grounded developer research assistant: it answers from data/corpus/, cites document ids, and refuses when nothing supports the claim. It is graded privately and earns the certificate. You build it in your own public repository: from the course folder, uv run bootcamp final new ../my-final-assignment, then cd ../my-final-assignment, uv sync, commit uv.lock, and gh repo create my-final-assignment --public --source . --push. Practise with uv run pytest and uv run bootcamp final grade, and hand in with uv run bootcamp final submit --github . Brief: units/en/unit2/capstone/. Every command, to your final score: the final assignment tutorial.

It is separate from the Gecko capstone, the store project presented on Friday 2 October, which lives in its own repository. Its code stays in your own my-gecko-buyer; you hand in a link to it by 5 October with uv run bootcamp gecko submit --repo ../my-gecko-buyer --github --push, or from the browser: hand in by 5 October.

A pass needs a score of at least 30% and every critical question. The final set has 15 questions, 6 of them critical, so any pass is at least 6/15 (40%). The starter refuses correctly and answers nothing: 3/10 (30%) on the practice set and 4/15 (27%) on the final set, failing both gates. Your latest submission counts. A pass earns an Ed25519-signed certificate anyone can verify. Details: final_assignment/README.md.

Bonus and extras

Optional. Never counted. Never required for the certificate.

• demos/ — notebooks run in class, to run again afterwards: an API request up close (real calls to a public pet API), one question three ways (prompt vs API vs MCP), regex, parsing and retrieval, Ollama on Colab for small laptops, the coach up close, Jupyter for beginners the coach in a chat, the weekly challenge, RAG on your laptop and your store, and a buyer.
• projects/ — real-world projects on open data with a local model and ChromaDB. Project 01: what are customers really saying? and Project 02: what are these companies worried about? (a RAG pipeline on real SEC filings, step by step, with a recorded lane for laptops without a model). Needs uv sync --extra projects.
• ship-it/ — turn the capstone into a surface somebody else can call, and a storefront an agent can buy from.
• Demos and explainers — Transformer Explainer, LLM Visualization, Tiktokenizer, Regexper, the Embedding Projector and more, each mapped to the session it helps.
• Two helpers inside every notebook: coach("…") answers from the course pages, and bootcamp_agent.patterns builds a regular expression from plain words — phrase("you must now"), near(one_of("send"), one_of("api key")) — so nobody has to write regex by hand.
• depth/ — six software-engineering modules (~45 minutes each): boundaries, data, architecture, reliability, production, graph RAG.
• cookbook/ — ten Gecko notebooks, from comprehending an OpenAPI spec to a verified loop on an instructor-hosted fork.
• workspaces/ — four open projects (corpus Q&A, retrieval lab, Autonomous Store, news-to-Telegram).

Commands

uv run bootcamp doctor                       # this machine
uv run bootcamp check ch03                   # one session scorecard
uv run bootcamp progress                     # every session, one tally
uv run bootcamp submit ch03 --github    # write submissions//ch03/

Checks judge behaviour, not wording. How to hand work in: units/en/unit0/how-to-submit.mdx. The generated index: docs/course-index.md.

Safety

Never place secrets in CLAUDE.md, AGENTS.md, Cursor rules, MCP JSON, issues, prompts, or commits. External-tool exercises use recorded/offline mode or the instructor-hosted fork. No wallets, no payment credentials, no production API keys, no mainnet path — ever — in class exercises.

Ask a coding assistant to inspect, propose a plan, and make the smallest change. Review the diff yourself. The loop lives in AGENTS.md.

License

MIT — see LICENSE.