Seven labs.
Real work, shipped.

The program Brief, with the room norms and what never goes in. The surprise: Claude is more capable than you think and less trustworthy than it sounds. A three-minute interview produces an instruction file that sounds like the organization, and a memo states, with confidence, things the emails never said. The person on the other end is the board member who reads that memo.

Claude interviews Chris for RW Institute on screen; he speaks his answers and Claude writes the AGENTS.md. Then a bare one-line request beside a brief in four plain parts (the goal and why it matters, the files to use, what done looks like, the limits), with the draft checked out loud against the practice email thread.

In rooms: eight minutes of introductions on "the task you'd hand off tomorrow," then 35 minutes of build: the model-training setting, the AGENTS.md interview, the readback test, and one small real item from each person's own notes or the practice email thread, checked by a fresh Claude session and then by the person. The outside read: Claude reads each person's AGENTS.md as a newcomer would, and the room hears one question from each.

The two questions, then each person writes a plan in the Zoom chat: "When ___ happens this week, I will ___." A participant gets the last word.

Three levels, each ending in something sent. Bring to Lab 2: the mission, an outcomes summary if there is one, and two writing samples (or a 10-minute voice note transcript with names cut, or the Lab 2 practice pack), a live RFP if there is one, and the workflow's name with its old-way time from a stopwatch.

The assumption on the table is "it can't sound like us." Two thank-you notes from the same one-line ask, one from a bare chat and one from a folder holding a fictional organization's writing, go side by side, and the room guesses which came from the folder. The person on the other end is the funder or donor who reads it.

The voice goes into files, on the practice pack: fill the Org-Brain folder, a short voice interview, one generic line cut from the voice notes by hand, a short draft from a brief with blanks, and a beat on keeping chats short to save usage.

Each person builds their own org brain from their own public documents or voice note, or the practice pack, whichever they pick, then writes one real piece or starts a compliance matrix on a live RFP. An outside read follows: a fresh Claude session lists the lines in the draft that don't sound like the voice notes, and each person reads aloud the line that sounds most like them.

The two questions, threads from the rooms, the plan in the Zoom chat.

Bring to Lab 3: one messy real reporting input, such as meeting notes, emails or a partial sheet, plus the name of the report and who reads it, or the practice mess pack.

The assumption is "if it reads well, it's right." The room votes on whether it would send a board update Claude drafted from the mess pack with one line of instruction, then hears what it got wrong and the file that shows it. The draft is re-tested on the current default model before the lab; if the model now handles the problems, the surprise becomes what it still misses. The person on the other end is the skeptical board member.

The second chair on the Brief's draft, on the practice route: a fresh chat, the second chair written live from the brief's questions, three findings settled (fix it, reject it with a reason, or check the source), then the second chair saved as a recipe with one catch added as a rule.

From this lab on, the brief comes as questions, answered in each person's own words. Everyone works solo on their own computer: each person takes one report section through draft, second chair and revision and saves the second chair in Recipes. Then each person gets an outside read: a fresh Claude session reads the revision as a different skeptic, a program officer or the board treasurer, and quotes the questions their second chair never asks, and each person reads aloud the one that matters most.

The two questions, threads from the rooms, the plan in the Zoom chat.

Each organization picks its Lab 4 material: its own recurring spreadsheet exported as a .csv, with every column that names or identifies a person deleted, staff included, and every notes column too, plus one total from its own system; or the practice sheet.

Two assumptions, both partly wrong: "our data can't go near AI" and "AI will just clean it." The people on the other end are the people inside the rows and the partners behind the placeholders, and the Brief is about why placeholders protect their dignity, the "With" posture.

The round trip on the practice sheet, 10 rows first: a copy with its headers checked, the key and the swap by hand, the brief with a change log, every change in those rows checked, the rest cleaned, the rows counted and the total added by hand, then the names back and the recipe saved.

Everyone works solo on their own computer, with their own Claude, copy and key. Red and notes columns are deleted by hand, then each person checks their own column headings, the top row only, with Claude before anything is saved. The placeholder key never goes near Claude or into the AI-Labs folder. In place of a swap, each person gets an outside read of their own cleanup recipe in a fresh session, then reads one rule aloud in a room round, placeholders only.

The two questions, threads from the rooms, the plan in the Zoom chat.

Bring to Lab 5: the meeting the next deck is for (who will be in the room, what they're deciding, and when) and its numbers with placeholders still in, or the practice set.

The assumption is "start in the slide tool." A deck made from the raw numbers says things the context would have stopped, and one made from a checked narrative holds together. The person on the other end is the meeting audience, who will check the math.

A checked narrative carried into a slide outline in a fresh chat, then the deck started. While it builds, a three-minute chart beat: a dishonest axis beside its honest twin. Then one number traced from a slide back to the narrative and the rows.

Rooms of about eight, from four organizations, each person on their own computer with their own Claude. On their own numbers, Claude interviews them for the context file; the practice set comes with one. Then insights, two checked by hand in the spreadsheet, the narrative, the second chair, the outline and the deck, each checked before the next. Near the end of room time, Claude plays the meeting's skeptic on one of their own slides in a fresh session, and in a room round each person reads aloud the question they'd least like to hear, 20 seconds each, with no answer.

The two questions, threads from the rooms, the plan in the Zoom chat.

On own numbers, a deck that isn't done gets finished at home, and the real names go back by hand in a copy outside the AI-Labs folder. Bring to Lab 6: a draft about to ship and the files it came from, and a research brief in the four parts, saved as Working/research-brief.md.

The assumption is "research gives answers." In a pre-run Research report, one confident claim's source is opened to show its date, then the sources are scrolled while the room names whose voices are missing. The people on the other end are the communities the sources miss.

The second chair turned into a skill on the practice route and run on a draft in a fresh chat, then where Research starts. Then five minutes on the tool map: the folder's AGENTS.md in a second tool with the readback test on screen, and Codex named in one line as a way to build a tool of your own, with no hands-on time.

Each person launches Research on their own question first, and it runs while they turn their own second chair into a skill and run it on a real draft. The brief is bare labels this week, only the four parts. An outside read follows: a fresh Claude session reads each person's skill and draft as a funder they've never applied to and names one question the skill never asks, and each person reads that question aloud in the room round.

The two questions, threads from the rooms, the plan in the Zoom chat.

The gather prompt fills the Workflows folder. Bring to Lab 7: the workflow's pieces gathered in Workflows (the recipes and skill it uses, and one recent output), and the ship log with its old-way and new-way minutes.

The facilitator reads back what participants said in Lab 1 about the task they'd hand off, in their own words. The assumption is "we learned some tools." The person on the other end is the colleague who will run the workflow.

Kit step 1 live: the brief, one interview answer, "done", and a walk through the saved workflow document.

The build and the showcase run in the usual breakout rooms of about eight, each facilitator taking the showcase notes in their own room. About 28 minutes to write the workflow document, get a cold read of it in a fresh session, and fix the gaps. Then the showcase in the same room: 90 seconds a person on what the workflow does, what it replaced, the old-way and new-way minutes, what surprised them and what it can't do, and one question from someone in another organization about what they'd need to run it from the document alone.

The program Debrief uses all three questions, adding "What might you do differently?"

A 90-day commitment and the last word to a participant. The 90-day plan takes the place of the homework levels, and month 1 is running the workflow for real. A separate showcase for Wells Fargo leadership and the nominators follows as its own session after Thanksgiving, in December or January, for a few standout projects. Its date isn't set yet.