This workshop is the fifth one this factory has built, and the first for an audience that isn’t already comfortable driving a coding agent. terminal-velocity, borrow-native, half-life, and heartbeat all assumed a practitioner who lives in git and a terminal. This one assumes someone who lives in Word, Excel, Outlook, and Teams, and has never touched either. That single change broke more of the existing pipeline’s assumptions than I expected going in.
The subject and the anchor
The subject is enterprise Microsoft 365 Copilot: Word, Excel, PowerPoint, Outlook, Teams, plus the Researcher and Analyst agents and in-app no-code agent creation. The curriculum anchor is AB-730, “Microsoft Certified: AI Business Professional,” a real, current, no-code Microsoft certification. Same pattern borrow-native used with Ardan Labs’ Rust certification: anchor to something real and proctored, and commit to personally sitting it once content exists, as dogfooding evidence rather than a marketing claim.
The gate problem this subject actually has
Every prior workshop had some kind of objective checker: a compiler and linter for Rust, a designed empirical experiment for the LLM-degradation subject, a designed network-fault simulator for distributed systems. A closed SaaS product has none of that. What we ended up with is a two-tier design: a required-output checklist (counts, exact-string matches, closed-set answers - checkable by counting or comparing, never by judgment) plus an AI-graded rubric, where the same product that produces the work also grades it. That second part is a real, named, unaddressed circularity, not something I’m pretending is solved. It’s in the design doc as an explicit hypothesis, not a settled fact.
A three-agent review chain, and what each stage actually caught
Before any of the phased build started, a plan for it went through Fable (a draft), Codex (an adversarial critique with live repo access), and a reconciliation pass that independently re-verified the disputed claims. Every single stage found something real the previous one got wrong. Codex caught that Fable’s exam-mechanics claims (a specific price and question count) were fabricated - neither holds up against Microsoft’s actual page. Codex’s own correction then invented a different, equally wrong price figure of its own, caught only when the reconciliation pass actually re-fetched the live page and found no price published at all. That’s the general shape of this whole session: a confident claim survives exactly until someone with tool access bothers to check it. (This paragraph deliberately doesn’t repeat any of those specific figures - see docs/brand.md’s exam-claim rule for why.)
Three rounds of doubt-driven development on the design doc itself
Once the design doc existed, it went through three adversarial review cycles (a fresh-context single model plus Codex cross-model, every cycle). Cycle 1 found the biggest structural problem: liability guardrails for the financial-services scenario were stated in one section and never actually wired into the modules most likely to violate them. Cycle 2 found that my own fix for cycle 1 introduced a real arithmetic bug (a coverage count that no longer summed to the right total). Cycle 3 found a Markdown table row I’d broken with an inserted blank line, which would have silently truncated the capstone’s own gate description on the published site.
The Review Panel’s most useful finding
Seven independent personas, one of them deliberately re-aimed from its usual “agent-literate practitioner” framing to this workshop’s actual audience. The single most severe finding: Module 03, the hard prerequisite for the entire back half of the arc, was gated almost entirely by the learner’s own unverified say-so that they’d saved and shared a prompt. Nothing proved the prompt was ever actually run against Copilot. Fixed with a real technique borrowed from Module 01: two actual transcripts, crafted prompt versus a naive one, diffed for a structural element the naive prompt’s real output doesn’t have.
The one part that was genuinely fun to build
The recurring scenario is Kellerman & Castle, a fictional financial-advisory firm whose entire cast is named after Patrick Swayze film characters, played completely straight through every exercise. It started as an internal joke and almost stayed permanently hidden - the workshop’s own Developer Evangelist review persona pointed out that was a wasted hook, since nothing about the actual grading mechanics depends on which names are attached to which role. coderturtle confirmed a one-time reveal after the capstone as the direction; the actual reveal mechanism (a footer credit, a Module 09 postscript, or a dedicated page) isn’t built yet.
What’s still open
Module content, fixtures, and grading prompts don’t exist yet - that’s Coachgremlin’s job, run later, one module at a time. The site you’re reading this on exists locally but isn’t deployed; GitHub Pages enablement and the first real workflow_dispatch are explicitly human-confirmed steps, not automatic ones. And there’s a real, unresolved maintainer question sitting in docs/next-actions.md: nothing yet confirms the person building this workshop actually has the enterprise Copilot tenant access the whole thing assumes.