Week 4 · Course 4

Advanced Workshop

From individual capability to organizational capability — map the frontier, ship a complex build, debug together, teach others.

Length4 hours (with one 10-min break)

FormatMicrosoft Teams workshop — bring builds in progress

PrereqAt least one deployed tool

Welcome — this is a workshop, not a lecture

How today is going to feel different from Weeks 1–3

What you do

Cameras on when you can
Have your tool open in another window
Bring at least one real problem to debug
Drop questions in chat live — we’ll surface them
Be ready to share screen on short notice

What I do

Talk less, watch more
Hold time on every activity
Surface patterns out loud as I see them
Refuse to fix your problem for you — coach you through it
Capture frontier findings to share back to the unit

You’re here because you’ve already done the work

Recap — what you brought into this room

Week 1

AI Fluency Fundamentals

The six 201 skills, the jagged frontier, centaur vs. cyborg, the management framing.

Week 2

Builder Orientation

Your first prototype. Task decomposition and iterative refinement applied to a real problem.

Week 3

Platform Training

Three complete tools on Power Platform, deliberately switching between centaur and cyborg work.

Required to be here: at least one tool you have actually deployed and somebody else is using.

What today is for

The shift from individual capability to organizational capability

You’ve been here

Built a tool
Made it work for you
Solved your own problem
Learned a few things the hard way

You’re going here

Map the frontier for your domain so others don’t fall in
Ship something complex on purpose, not by accident
Debug other people’s tools, not just your own
Build a QA habit you can hand to a junior Marine
Teach the 201 skills forward

Agenda — 4 hours

Workshop blocks are gold-coded throughout the deck

Time	Module	Duration	Mode
0:00–0:30	M1 — Frontier mapping for your domain	30 min	Workshop
0:30–1:30	M2 — Complex build (Unit Readiness Dashboard)	60 min	Live build
1:30–1:40	Break	10 min	—
1:40–2:20	M3 — Group debugging (real problems)	40 min	Clinic
2:20–2:50	M4 — Verification protocols and QA	30 min	Reference + drill
2:50–3:20	M5 — Teaching others (teach-back)	30 min	Workshop
3:20–3:50	M6 — Workflow playbook	30 min	Workshop
3:50–4:00	Wrap-up, certification, Week 5 preview	10 min	—

Module 1

Frontier mapping for your domain

30 minutes

Make the boundary between “AI helps” and “AI hurts” visible — for your specific work, in writing, shareable.

Open instructor notes for this section ↗

Why this map is the most valuable thing you’ll make today

−19pp

Performance drop when consultants used AI on tasks outside the frontier — vs. working without AI at all.

BCG · Harvard, 758 consultants, 2023

What that means today

The boundary is real, and it’s not where intuition puts it.
People don’t notice when they cross it.
Quality collapses silently — the output still looks right.
The map is what makes the boundary visible to your section.
Most valuable artifact you’ll leave with today.

Worked example — 1st Bn, 99th Marines

A frontier map looks like this. Yours will look like this.

Category	Inside frontier (AI handles)	Outside frontier (AI fails)	Moving frontier (re-test)
Document generation	Counseling drafts, award write-ups, memo formatting	Local SOPs, exact regulation quotes, unit-specific policy	Fitness report narratives, legal review summaries
Data analysis	Trend ID in clean data, dataset summaries, anomaly flagging	Operational interpretation (“why did readiness drop?”)	Predictive analysis — retention, maintenance forecasting
Process automation	Approval routing, notifications, status tracking	Multi-system integrations with legacy DBs, judgment calls (hardship)	Complex conditional workflows with clear business rules
Reference lookup	Finding MCOs/NAVMCs, summarizing policy docs	How a regulation applies to this edge case	Authoritative interpretation — still unreliable
Training development	Lesson outlines, quiz generation, slide structure	Effectiveness evaluation, MOS-specific accuracy	Full lesson plan generation — quality varies

What good looks like

Before you start typing — here’s the bar

Specific, not generic

Bad: “writing”
Good: “counseling statement drafts for Page 11 entries”
Each cell is a real task somebody on your team does this month.

Tested, not guessed

Every “outside frontier” entry should be something you actually saw fail.
If your “outside” column is empty, you’re overestimating AI.

Per tool, not in general

GenAI.mil and CamoGPT have different frontiers in your domain.
Note which tool the entry applies to.

Living, not finished

“Moving frontier” entries get re-tested next quarter.
Date the map. Initial it. Hand it off when you PCS.

Workshop · Build now

Build your frontier map — right now

15 minutes · silent work, then share-out

Prompt Open a blank doc. Make four columns: Category · Inside · Outside · Moving. Pick five categories that cover what your section actually does this month. Fill at least one specific, real example in every cell — tasks you have personally seen succeed or fail.

What good looks like in 15 minutes: 5 rows, every cell has at least one specific task (not a topic), and your “outside frontier” column is not empty. If you’re stuck, look at your last 30 days of work and ask: where did AI save me time, and where did it cost me time?

Share-out — round the room

10 minutes · 90 seconds each · we are listening for patterns

What to share

One inside-frontier task that surprised you (AI is better at this than you thought).
One outside-frontier task you have personally seen fail.
One thing you’d add to the “moving frontier” column that you want to re-check next quarter.

What I’m capturing in chat

Failure cases — these become unit-wide frontier intel.
Cross-tool differences (GenAI.mil vs. CamoGPT).
Anything 2+ people independently discovered.

Hard stop at 10 minutes. If we don’t get to you, drop your map in chat.

Module 1 — takeaway

The frontier map is the artifact that prevents the 19-point performance drop. When you can name where AI fails in your domain, you stop trusting it there — and you stop the rest of your section from trusting it there too. Keep your map. Date it. Share it back to your unit by EOW.

Coming up next: a build hard enough that you’ll have to deliberately switch modes three times.

Module 2

Complex build — multi-component system

60 minutes

Unit Readiness Dashboard. Three phases. Three deliberate mode switches. The point is the decision-making, not the polish.

Open instructor notes for this section ↗

Before we touch the keyboard

The thing I’m grading is your mode-switching, not your dashboard

Mode-switching is the goal. Watch for the moments where you should slow down for accuracy and the moments where you should iterate fast. That’s the skill.

Centaur — you frame, AI drafts, you verify before moving on. Use it when errors compound.
Cyborg — rapid back-and-forth, fail fast, feed errors back in. Use it when iteration is cheaper than verification.
The trap: staying in cyborg mode for accuracy-critical work because it feels productive. It isn’t.

Build goal — Unit Readiness Dashboard

Power BI · pulls from three data sources · ~10 prompts · ~45 minutes

Inputs

Training tracker (Excel)
Equipment status (CSV from GCSS-MC, or simulated)
Personnel roster (Alpha roster, or simulated)

Outputs

Overall readiness % (color-coded card)
Bar chart by company (Alpha, Bravo, Charlie, Weapons)
Table of personnel not ready

You will switch modes 3 times

Phase 1 — Centaur (data architecture)
Phase 2 — Cyborg (data ingestion)
Phase 3 — Centaur (visualization)

If you finish, the dashboard is a bonus. The grade is the mode-switching.

Three phases — map your mode switches

Phase 1 · 15 min

Data architecture

Centaur

Why: schema errors compound. Verify before building.

Prompt for star schema
Verify columns exist in real data
Decide: rename source, or map in Power Query?

Checkpoint: schema matches your actual columns.

Phase 2 · 15 min

Data ingestion

Cyborg

Why: things will break. Feed errors back in.

Generate M code per source
It will fail — that’s expected
Paste error + actual file structure back in

Checkpoint: all three sources load clean.

Phase 3 · 20 min

Visualization

Centaur

Why: leadership briefs from these numbers.

DAX measure for readiness %
Hand-calc one company — numbers must match
Spot-check one Marine in the “not ready” list

Checkpoint: every number traces to source data.

Switch · Live build

Open Power BI · share screen if asked

You have 45 minutes. I will walk the room.

Type your phase number in chat when you switch modes — P1 · P2 · P3.

If you finish early: help your neighbor. Don’t polish.

Welcome back — debrief

10 minutes · volunteers, then anyone I haven’t heard from yet

Where did you switch modes? Why there?
Where did the AI fail? Was it a frontier issue, a context issue, or a platform quirk?
If you had to build this again tomorrow, what would you do differently?
How long did it take you? How long would it have taken without AI?

10 prompts. 3 mode switches. At least 2 error-recovery cycles. That’s normal. The students who finished are the ones who verified at each phase boundary and fed errors back in.

Module 2 — takeaway

Complex builds aren’t harder because the prompts are harder. They’re harder because you have to keep changing how you work with the AI. The skill we’re grading is noticing the moment you should switch modes — and actually switching.

Time for a 10-minute break. We come back at the time on the clock, not the time you sit down.

BREAK

10 minutes · back at the time on screen

Module 3

Group debugging — real problems

40 minutes

You brought broken tools. We’re going to fix the diagnostic pattern, not the tool. 5 problems · 7 minutes each.

Open instructor notes for this section ↗

Debugging clinic protocol

Same four steps for every problem · 7 minutes total · this slide stays up

Step 1 · 2 min

Student presents

Format: “Expected behavior … Actual behavior … Steps I’ve already taken.” No story-telling. No theories. Just those three lines.

Step 2 · 3 min

Group diagnosis

Group asks clarifying questions and proposes hypotheses. I’ll prompt: data or logic? input or output? Have we seen this pattern before?

Step 3 · 2 min

Instructor synthesis

I name the root cause category — frontier limitation, missing context, wrong assumption, integration failure, data quality — and how to approach the fix.

Step 4 · in chat

Document the pattern

The failure case goes on the collective frontier map — even if we didn’t fully fix it. The pattern is the deliverable, not the fix.

Facilitation rules — for the room and for me

When someone shares a real problem · how to keep it inclusive · when to cut it

Make it safe to share

Praise the broken tool out loud — this is the most useful thing you brought.
Never let “you should have known” into the chat.
Reframe blame as pattern: “That’s a context-gap pattern, not a you-pattern.”
Ask the presenter what kind of help they want before the group jumps in.

Keep it inclusive

Call on different people each round — not just the loudest hand.
Quiet folks: invite them by name for clarifying questions, not solutions.
Run hypotheses in chat in parallel with voice — gives introverts a lane.
Two-question max per group member per problem.

When to move on

Hard stop at 7 minutes. Set a visible timer.
If we name the root-cause category, that’s a win — even unfixed.
If the group is rabbit-holing on syntax, I cut it: “Pattern named, post fix in chat after class.”
Aim for 5 problems in 35 minutes. Save 5 minutes for synthesis.

Backup scenarios — if the room is light on real problems

Use one of these to fill the queue. Full write-ups in the instructor guide.

Scenario A — Stale data in flow

Power Automate fires on update, but the email always has the old values.

Pattern: platform timing quirk — not a frontier issue.

Scenario B — Form saves invalid data

Power Apps phone-number validation “works,” but the submit button still accepts “abc123.”

Pattern: context gap — AI wasn’t told the button must respect the validation state.

Scenario C — Dashboard shows stale data

Power BI scheduled refresh “completes successfully” but nothing updates.

Pattern: domain knowledge — data source pointed at a local C:\ path, not SharePoint URL.

Notice the through-line

All three look like AI failures. None of them are. They’re context gaps and platform quirks.

That’s the pattern most debugging surfaces.

Synthesis — what did the room just learn?

5 minutes · group discussion before we move on

What patterns repeated? (Most problems are context gaps, not frontier limits.)
How many were “AI couldn’t” vs. “we didn’t tell it enough”?
What questions should we ask the AI differently next time?
Which problems belong on the frontier map?

Most debugging is one of three things: the AI didn’t have enough context, the platform has a quirk the AI doesn’t know about, or we hit a real frontier. Categories 1 and 2 we can fix. Category 3 goes on the map.

Module 4

Verification protocols and QA

30 minutes

A reference card you can hand to a junior Marine, plus a timed drill that proves you can spot what AI gets wrong.

Open instructor notes for this section ↗

QA reference card — the five protocols

This slide stays on screen for the discussion. Screenshot it.

Protocol 1

Source verification

AI fabricates references. Every citation, regulation number, NAVMC form, and URL is independently verified against the official publication system. If it can’t be found, it’s wrong.

Protocol 2

Data accuracy

Every number, date, name, rank, and quantity gets checked against source data. Spot-check at minimum. AI confidently uses plausible-but-wrong figures.

Protocol 3

Logic check

Does the reasoning hold end to end? Are conclusions supported by the premises? Watch for steps in wrong order, contradictory timelines, and silent assumptions.

Protocol 4

Format compliance

Does the output match required formats, templates, DTGs, and standards? AI loses formatting consistency in long documents — especially in numbered lists and procedures.

Protocol 5 · the one nobody can skip

Domain review

Does this pass the smell test for someone who actually does this work? Hand it to the SME before you sign it. The other four protocols are mechanical — this one is judgment, and it’s the one AI can’t do for you.

Workshop · Timed drill

Find the five errors

10 minutes · silent · then 10-minute debrief

Prompt I’m sharing a one-page AI-generated SOP excerpt in chat now. Run it through all five protocols. Mark every issue you would not sign. Number them. There are five planted errors: two fabricated references, one contradictory timeline, one logic error, one format break. Find all five.

What good looks like: all five errors found in under 10 minutes, with the protocol number that caught each one. If you only find three, you’re skipping a protocol — usually source verification, because it requires looking things up. Looking things up is the protocol.

Module 5

Teaching others — the 201 multiplier

30 minutes

You don’t leave this room as a graduate. You leave as the person who teaches the next two Marines what you just learned.

Open instructor notes for this section ↗

What you owe forward

The 201 multiplier · ties to the EDD train-the-trainer guidance

The owe

You teach two Marines the 201 skills before your next PCS.
You sit in on someone else’s session and give one piece of feedback.
You contribute one entry to the unit’s frontier map per quarter.
When you PCS, your replacement is certified before you walk.

Why it’s on you

Mollick: workers are using AI but hiding it. Shadow culture, no shared practice.
Your job is to surface it, formalize it, and pass it on.
Each unit needs 2 certified instructors at all times — succession is your responsibility.
Train-the-trainer pathway is in the EDD guidance: shadow → co-teach → deliver.

The apprentice problem — and the four protocols

Entry-level postings dropped 35% in AI-exposed jobs from 2023–2025. If juniors never write the doc, how do they learn to spot when AI’s draft is wrong?

What you do with juniors

Require review and explanation — they have to defend why AI’s output is right or wrong.
Periodically work without AI — key tasks done from scratch monthly to keep the muscle.
Use AI output as a teaching tool — hand them a flawed draft, have them find the issues.
Rotate them through QA — the only way they build judgment is exposure to bad output.

What you don’t do

Skip the apprentice phase because AI is faster.
Let juniors paste output directly into final products.
Use AI to grade AI without a human in the loop.
Assume the section will be fine if you stop developing the next generation.

Workshop · Teach-back

Teach one concept — 3 minutes

2 min pick · 5 min prep · 10 min round-robin in breakouts · 5 min debrief

Prompt Pick one concept from EDD: centaur vs. cyborg, frontier mapping, context-building, iterative refinement, verification protocols, the jagged frontier. Use the template: one-sentence definition · why it matters · one real example from your job · one common mistake · one takeaway. Then teach it in 3 minutes in your breakout room.

Peers grade you on three things: could they explain it back to someone else, was your example concrete enough to be believable, and do they know what to do differently tomorrow because of you. Time-box hard. Brevity is the teaching skill.

Module 6

Workflow playbook & wrap-up

30 minutes

One page. One recurring task. One AI-integrated workflow your section can run without you. The deliverable that proves you graduated.

Open instructor notes for this section ↗

What a finished playbook looks like

Worked example · weekly training schedule publication

Task	Weekly training schedule for the section
Frequency & mode	Every Thursday by 1600 · Cyborg (continuous back-and-forth)
Steps (Human / AI)	1. H: Pull events from calendar, OPORD, taskings. 2. AI: Draft schedule in standard weekly format. 3. H: Cross-reference range bookings, vehicles, instructor availability. 4. AI: Format conflicts as a decision matrix with options. 5. H: Decide conflicts, add section leader notes. 6. AI: Generate the final formatted schedule for distribution.
Verification checklist	Every event has a confirmed location · all times in 24-hour · no double-bookings · uniform specified · POC listed.
Known frontier issues	AI invents room numbers · cannot check range availability · sometimes reverts to 12-hour time.
Time savings	~3 hours without AI → ~45 minutes with AI.
Junior development	Rotate schedule duty among juniors weekly. They brief why each event is scheduled. Once a month: schedule built without AI to keep the baseline skill.

Workshop · Build now

Write your playbook — one page, one workflow

20 minutes · silent work · submit in chat by the timer

Prompt Pick a real, recurring task from your job. Fill the seven fields: task · frequency & mode · 4–8 H/AI steps · verification checklist (3–5 items) · known frontier issues · time savings · junior development note. It must be specific enough that a junior Marine could run it without you in the room.

Completion bar: 4+ steps, each labeled H or AI, a verification checklist with 3+ items, at least one real frontier issue you’ve hit, a time savings estimate from your own data, and a junior development note that’s actionable. If your playbook has no verification checklist, it isn’t done.

Reflection — before you close the laptop

3 prompts · drop one-line answers in chat

What’s the one thing from today you’re going to do differently next week?
Which of the six 201 skills did you feel weakest on today — and what’s your plan?
Who’s the first person you’re going to teach this to, and when?

I’ll save the chat and send the synthesis to your section lead.

Certification path — what it takes

Meets-or-better in 5 of 6 categories · recommended for instructor track

The six rubric categories

Frontier map completeness (M1)
Complex build & mode-switching (M2)
Debugging contribution (M3)
QA protocol rigor (M4)
Teaching effectiveness (M5)
Workflow playbook completeness (M6)

What “graduated” unlocks

Serve as a Platform Training instructor.
Lead tool-development projects in your unit.
Mentor junior personnel in AI-assisted workflows.
Contribute to frontier map and workflow playbook libraries.
Path to QA reviewer → Advanced Workshop instructor.

Didn’t hit the bar today? Build 1–2 more tools, attempt Advanced Workshop again. We’ll keep the seat.

Next week — Week 5: Supervisor Orientation

30 minutes · for leadership · you should encourage your chain to take it

What it covers

How to evaluate AI tool proposals from your Marines.
How to create a permission culture so people stop hiding their AI use.
The apprentice problem — from the supervisor’s seat.
What to ask in a 1-on-1 about AI work.

Your action item

Forward the Week 5 invite to your section leadership today.
Bring your frontier map to your next 1-on-1 with your supervisor.
Submit your playbook to the unit registry by EOW.

Thank you. Now go teach somebody.