Skip to content

How the Lab Works

This lab is built around Claude Code as your AI development partner. You don't copy-paste commands from a PDF -- you work interactively with an AI mentor that knows the codebase, paces its teaching, and adapts to your progress.

Claude Code as Your Mentor

Claude has been configured (via CLAUDE.md in the repo root) specifically for this lab. It will:

  • Teach one concept at a time and check your understanding before moving on
  • Ask questions, not lecture -- Socratic method throughout
  • Run commands for you and explain what happened -- you learn the reasoning, not the syntax
  • Give progressive hints when you're stuck:
    1. Concept hint -- what you should be thinking about
    2. Approach hint -- how to get there
    3. Specific hint -- exactly what to do
  • Show real code from the project instead of abstract examples

Language Support

Claude supports the lab in your native language. If you write in Spanish, Portuguese, or another language, Claude will respond in that language automatically. Technical terms (pipeline, endpoint, scan, AIRS, IAM, etc.) stay in English since they're industry-standard. Quiz answers are evaluated on concepts, not language — answer in whatever you're most comfortable with.

Experimental

Multilingual support is built in but not yet extensively tested across all modules. English is the primary language. If you hit rough edges in another language, let your instructor know.

Commands

Everything runs through Claude Code slash commands:

CommandPurpose
/lab:module NStart or resume module N
/lab:verify-NRun verification checks for module N
/lab:hintGet a progressive hint for your current challenge
/lab:explore TOPICGuided deep-dive into a concept
/lab:quizTest your understanding
/lab:progressSee your completion dashboard

Scoring & Leaderboard

Your progress is tracked and scored throughout the lab. Points come from two sources:

Technical Checks

Each module has concrete verification steps that Claude runs against your actual infrastructure. These check things like: Is your GCP project configured? Did the training job produce output? Is your app deployed and responding? You earn points when real checks pass -- not self-reported.

Understanding Questions

After completing a module's challenges, Claude asks comprehension questions to test your understanding. These are scored 0-3 per question:

ScoreMeaning
3Complete, accurate answer
2Mostly correct, minor gaps
1Needed hints to get there
0Could not answer

Collaboration Bonuses

Instructors can award bonus points during discussion breaks:

BonusPointsHow
Teaching Bonus+2Explained a concept to a classmate
Discovery Bonus+2Found an undocumented issue or insight
Best Question+1Asked a particularly insightful question

Points by Module

ModuleTechnicalUnderstandingMax
0: Environment Setup14822
1: ML Fundamentals--99
2: Train Your Model268
3: Deploy & Serve4610
4: AIRS Deep Dive8614
5: Pipeline Integration10313
6: The Threat Zoo6612
7: Gaps & Poisoning6612
Total5050100

Workshop students

Workshop scenarios include 2 additional points in Module 0 (upstream remote setup and project naming verification) for a possible 100-point maximum. Self-paced scenarios cap at 98.

Leaderboard

After each /lab:verify-N, your score is automatically posted to the live leaderboard. You can see where you stand relative to other participants in real time.

The leaderboard tracks:

  • Total points across all modules
  • Number of modules completed
  • Individual module scores

Pacing & Hard Stops

Workshop Pacing

The lab is designed for ~4 hours of active work, split across two sessions with a presentation break in between.

Hard Stops

In workshop scenarios, there are hard stops after Module 3 and optionally after Module 4. When you hit a hard stop, Claude will tell you. You cannot proceed to the next module until the instructor resumes the group.

Why? The instructor-led AIRS presentation between Acts 1 and 2 builds on what you just experienced. If half the class has already raced ahead, the discussion loses context and everyone suffers.

If You Finish Early

Help others. This is collaborative, not just competitive. The Teaching Bonus (+2 pts) and Discovery Bonus (+2 pts) reward exactly this behavior.

Walk over to someone who's stuck. Explain what you figured out. You'll solidify your own understanding and earn bonus points.

If You're Falling Behind

The leaderboard updates live. If you notice you're one of the last to complete a module, pick up the pace. It's fine to dig deep into interesting topics with Claude -- that's great learning -- but be mindful of the group's timing.

Time budgets

ActModulesTarget Time
Act 1: Build It0, 1, 2, 3~3-4 hours
Presentation Break--~30-45 min
Act 2: Understand Security4~1-1.5 hours
Act 3: Secure It5, 6, 7~2.5-3.5 hours

Hard Blockers

Some setup issues in Module 0 are hard blockers -- if your GCP project isn't accessible or AIRS credentials aren't configured, later modules will be blocked. /lab:verify-0 catches these early. Don't skip it.

If you hit a hard blocker, Claude will warn you clearly and add it to your progress file. You can still participate in discussions and concept exploration, but the technical challenges that depend on the blocked resource won't work.


Scenarios

When you first start the lab, Claude asks which scenario you're in:

ScenarioHard StopsLeaderboardFor
TS WorkshopYesYesInstructor-led Technical Services workshop
TS Self-PacedNoYesSelf-paced TS learning
InternalNoNoOther internal teams
PublicNoNoSelf-guided learners

Your scenario determines pacing rules, whether hard stops are enforced, and whether your scores post to the leaderboard. Claude handles all of this automatically based on your selection during onboarding.

Built with Claude Code