Seven labs.
Real work, shipped.

Agency framing, not a tool pitch. The fears in the room (accuracy, voice, data, jobs) get named directly, each one mapped to the specific practice that addresses it.

The round-trip end to end on a mock email-thread export: swap names for tokens, drop the file into the working folder, draft a board update with a narrated prompt, catch a planted contradiction in verification, return and re-identify. That motion is the whole program.

Pairs run the identical exercise on the mock pack first (completion, not quality), then aim the same loop at the smallest writing item on their own ship list.

Two or three pairs show what they made, struggles included. Each pair states its homework commitment out loud.

All three levels run the round-trip on real work in their own environment; everyone also gathers 3 to 4 public org documents for next week's build.

Homework debrief: two pairs show the writing they shipped, including what broke.

An org brain built live: a Project loaded with mission, writing samples, and outcomes, then voice instructions written by describing the samples and editing by hand. Payoff: a mock RFP becomes a compliance matrix and a first draft. The same prompt runs once without the Project; the side-by-side makes the argument.

Pairs build their own org brain from material they brought (a fictional starter pack covers anyone who arrived empty-handed), then aim it at a writing item from their ship list. Floor: Project exists, one draft out. Ceiling: voice tuned through two revisions.

Before/after read alouds. Where did it still not sound like you, and what instruction fixes it?

One more writing rep, a full shipped deliverable, or a second voice profile (funder-formal vs. community newsletter).

Homework debrief, plus one question: did anyone's Project surprise them?

Real reporting inputs (meeting notes, a partial spreadsheet, three emails) become a plausible board summary. Then the move this lab exists for: a second prompt played as a hostile board member attacks the draft and catches planted weaknesses on screen, including an unsourced statistic and the sentence that would hurt if quoted out of context. One review, two protections: accuracy and communication risk. Plausible is the danger zone.

Pairs run the same arc on the mock pack, then on a real reporting item. The adversarial prompt and a source-accountability checklist get saved into every Project before the lab ends.

Each pair shares the single best catch its critique made. Best catches are the proof the step earns its time.

Critique something already in your outbox, ship a full report section through the arc, or tune the critique into your org's own failure-mode profile (the seed of the Lab 6 agent).

The week's best adversarial catches, fast.

A 58-row mock grants export carrying every real disease: one org spelled three ways, blanks, amounts as text, three date formats, a duplicate. First a green/yellow/red classification of what can travel and how, then two separate passes in Cowork: clean with a row-by-row change log, then insights (totals, outliers, what looks wrong). Verification is specific: spot-check five rows, reconcile totals, calculator-test the findings. The change log is the star, not the clean file.

Pairs run the mock dataset through the full arc, then write their own cleanup recipe: a saved, named prompt for the recurring mess they actually get every month.

Recipes read aloud; the room steals freely. The same five messes live in every org.

Run the recipe on real data, deliver a full clean-plus-insights memo, or chain recipe into a standing monthly review.

One real cleaned-data win from the homework, told by the pair that had it.

Three moves on the cleaned data. Numbers to narrative: an impact story for a named audience where every figure traces to a source row, then an adversarial pass played as a funder who suspects cherry-picking. Visual judgment: chart or sentence, one message per chart, honest axes, and Claude proposes the chart while the human edits the judgment. Narrative to deck: slide titles as claims, numbers in speaker notes, generated in minutes. The deck is fast because the thinking was already done.

Pairs work the ladder on their own Lab 4 output (or the reference data): narrative first, deck outline second, one chart specified per the visual guide.

Two pairs present one slide each, 60 seconds, as if to their real audience. What would the hostile funder ask?

Finish the deck (the second rep that makes it stick), build the full pack for a real upcoming meeting, or template a reusable board-pack generator.

One deck from the homework, 60 seconds.

Two moves. The critique prompt becomes a standing agent: named, saved, triggered on demand, with org-specific failure modes. A prompt is a thing you remember to do; an agent is a thing that happens. Then the wider toolbox, demonstrated live by the facilitator: an agentic research run (a funder-landscape scan) launches while the model-selection map gets drawn (Copilot for inside the container, Claude for the work, research agents for the outside world, NotebookLM for your own documents and notes). The scan returns and gets verified live, with the bias question asked out loud: whose perspective did these sources systematically miss?

Pairs build and test their critique agent on a Lab 5 artifact, writing failure modes from real experience (what does YOUR board push on?). The research motion runs on the facilitator's screen with pairs directing it, unless Wells Fargo opts into participant accounts; three claims get verified before anything is repeated.

Each pair states one sentence: the next thing we'd automate and the tool we'd use. Those sentences come back in Lab 7.

Use the agent on a real deliverable, write up a verified research answer, or build a second agent or prospect pipeline (the champion track that shapes what comes after the pilot).

The frame shifts from learning to owning. The automation sentences from Lab 6 get read back, verbatim.

One complete workflow document, filled in honestly: trigger, steps, the Project artifacts that make it run, verification gates, and real before/after time costs. A workflow nobody wrote down is a demo, not a workflow.

Assembly, not construction: pairs document the workflow from pieces already built across the homework arc (org brain, critique agent, cleanup recipe, narrative-and-deck motion). The floor version, round-trip plus one motion, is a legitimate, complete outcome. One test: could a colleague run this from the document alone?

Two minutes per pair, hard-timed, four beats: what we built, what it replaced, what it costs us now, one number. Everyone else writes one steal per presentation. No critique today; two minutes of being proud of work is part of the design.

90-day commitments written and spoken: the workflow runs at least monthly, one optional Build Ahead item, and where help lives when something breaks. No homework; the 90-day plan is the homework.