Most winery owners I talk with want to know which AI tool to buy, which prompt library to copy, which 30-day plan to follow. The single highest-leverage move in the whole rollout is none of those. It is 90 minutes, in person, with one employee, before day one of the play starts.
Those 90 minutes are when you and your first AI champion build her starter folder together. A short role description in her own words. A voice file with 5 to 10 examples of how she writes on the job. A description of the recurring task you have picked. Three sample inputs. Three sample outputs. That is the folder she opens on day one. Without it, day one is "go figure it out," which is the failure mode for most adoption attempts I see.
I have come to think of this 90 minutes as the most important time a leader spends in the entire 60-day play. The hours that follow — picking the tool, scheduling the Friday ritual, deciding whether to use Claude or ChatGPT — are real work. They are also reversible. If you get any of them wrong in week 1, you fix them in week 2 and the rollout keeps moving. The starter folder is the one piece that, if you skip it, the rollout never gets to week 2.
Why day one collapses without it
A blank screen on day one quietly ends the rollout before week two.
When an employee opens a fresh ChatGPT or Claude window with no setup, she faces the same cognitive hurdle a new writer faces with a blank page. She does not know what tone the tool should use, and she does not know whether the result will sound anything like her or like a generic AI assistant pasted into her job. Most people in that position do one of two things. They type a short generic prompt, get a generic response, decide AI is not useful for their work, and close the tab. Or they get something close-but-wrong, spend twenty minutes editing it back into their voice, and conclude the tool is slower than just writing the email themselves.
Either path ends the same way. By Friday of week one she has opened the tool twice, gotten two unsatisfying results, and stopped opening it. Day 30 telemetry shows zero use. The leader concludes she picked the wrong employee or the wrong tool. The cause is usually neither. The setup was the wrong shape.
A starter folder turns day one from "go figure it out" into "use this." The role description tells the tool who she is. The voice file teaches it her cadence and word choices. The task description sets the bar. The sample inputs and outputs are a working demonstration of what good looks like. She does not need to invent any of that on her first session. She pastes a real input into a prompt that references the folder, the tool produces a draft that sounds roughly like her, and she edits the last 20 percent instead of writing the whole thing.
The first session ends with a usable artifact. That is the entire ballgame on day one.
What the folder is built around
A starter folder is built around exactly one thing: the hated task your champion already complains about every week.
If you have not picked that task yet, the 90 minutes does not happen. There is nothing to anchor the role description against, nothing to populate the sample inputs with, no clear definition of what success looks like for the sample outputs. Leaders who try to build a starter folder around a vague aspiration — "she'll use AI for marketing stuff" — produce a folder that fits no real Monday-morning workflow. It sits unused.
Pick the task first. Tasting room visit recaps, club cancellation responses, wholesaler follow-ups, monthly newsletter drafts, harvest paperwork. The kind of work that shows up every week with real artifacts attached and low stakes if a draft misses. Then the 90-minute build has a target.
What goes in the folder, and why it goes there
The folder has five elements. Each one solves a specific failure mode I see in rollouts that skip this step.
The role description is two to three sentences. Who she is at the winery. What she is responsible for. What tone she uses with customers. Not a job title. A working description, in her own words, that the tool can read as a system prompt. This solves the problem of the tool defaulting to a generic professional voice that nobody at the winery would write in.
The voice file is 5 to 10 examples of her actual writing. Three real customer emails she is proud of. Two internal notes. A thank-you note she wrote last month. Labeled at the top, "this is how I write." The voice file is the single most underrated artifact in the whole rollout. Without it, every output reads like generic AI prose. With it, the tool produces drafts that sound enough like her that the editing pass takes minutes instead of forty.
The task description says what she is trying to produce, who reads it, what good looks like, what off looks like. Two paragraphs is plenty. The "what off looks like" part is the one most leaders skip and the one that does the most work. A model with no negative examples produces overconfident, slightly wrong drafts. A model that knows "do not use the word 'exquisite,' do not call our wines 'world-class,' do not open with the customer's first name in all caps" produces drafts she can ship.
The three sample inputs are real examples of what she starts with on a Monday morning. A customer email that needs a reply. A tasting room visit she needs to write up. An order she needs to confirm. Pulled from her actual inbox or her actual ticket queue, not invented. The reason these have to be real is the same reason the voice file has to be real: the tool has to learn the shape of her actual work, not a sanitized version of it.
The three sample outputs are what she would ideally produce from each input. Even if rough. Especially if rough. These show the tool the gap between the raw input and the finished artifact. With three before-and-after pairs in the folder, a model can pattern-match its way to a usable first draft within a few prompts. Without them, the tool guesses.
That is the whole folder. Five elements. One Word doc or one shared note or one folder of five files, however your champion likes to organize her work. The format does not matter. The content matters.
Why it has to be in person, and why it has to be 90 minutes
I get asked some version of "can we do this asynchronously?" every time. The answer is no, and the reason has nothing to do with productivity theater.
The 90 minutes is a working session, not a briefing. The two of you sit at the same screen, pull real examples out of her sent folder, talk about which ones sound most like her, debate what "off" looks like for her department, and produce a folder that is hers. The conversation that happens while you are doing it is half the value. She gets to tell you which task drains her week (the one you think drains her week is often a different task). You get to hear which words she would never say. The folder ends up shaped by her, which is what makes her open it on day one.
If you do this over Slack or by sending her a template to fill in on her own, you get a folder that looks right and works in a pilot demo and falls apart on contact with her real Monday. She has no investment in it. It is your folder, not hers, and on day one she opens it once, finds it generic, closes it, and goes back to writing the responses by hand.
90 minutes is the budget. Not 20, not 4 hours. 20 is not enough to pull real examples and have the conversation. 4 hours becomes a workshop, and workshops end without artifacts. 90 minutes is long enough to get the folder built and short enough that both of you protect it on your calendars. If you finish in 75, fine. If you run to 110, fine. Block 90.
It does not have to be perfect. Version one of the folder will be wrong in places. The voice file will be missing examples. The sample outputs will have rough edges. That is intentional. She will fix it in weeks 2 and 3 once she sees how the tool uses each piece. The point of the 90 minutes is to leave the room with a folder she can open on day one.
A concrete example: a wine club concierge
To make this concrete without inventing real-winery facts, here is what the build looks like for Not Really Wines, our fictional demo winery. Imagine Priya Sandhu, the DTC director at NRW, has picked her wine club concierge as the first AI champion. The hated task is club cancellation responses — a dozen a week, each one taking 15 minutes, each one needing the same warm-but-precise tone that NRW's brand voice calls for.
Priya blocks 90 minutes on Wednesday afternoon. She sits down with the concierge with two laptops and a shared doc open.
They write the role description first, in the concierge's own words. "I respond to wine club members who are pausing or cancelling their membership. My job is to make them feel respected, offer one real alternative if it fits, and keep the door open if they ever want to come back. Members read my emails and decide whether to recommend NRW to their friends."
They pull five real cancellation replies the concierge has sent in the last two months. They pull two thank-you notes she wrote to members who stayed after a billing problem. They paste all seven into a "voice file" section labeled "this is how I write." Priya notices that the concierge uses "while you're away" instead of "during your pause" and writes that down in the task description as a voice tic to preserve.
They write the task description together. What good looks like: warm, three sentences max, one specific alternative offered if relevant, no upsell pressure. What off looks like: anything that reads like a retention script or opens with "I'm sorry to hear that," and any mention of percentages or comp offers the member did not ask about.
They pull three real cancellation emails from this week's queue — sanitized of member names — as the sample inputs. They paste the responses the concierge already drafted as the sample outputs. Three input-output pairs, in the folder, in the concierge's own voice.
The folder is done in 80 minutes. The concierge opens it Monday at 9 a.m. She runs the first cancellation reply through Claude using the folder as context. The draft comes back sounding close enough to her that she edits two phrases and sends it. The reply went out in 4 minutes instead of 15.
That is what day one looks like when the 90 minutes happened.
What this means for what you do next week
If you are about to start an AI rollout for your winery and you have not yet picked the champion and the task or built the starter folder, those are the pieces to put in place before day one of the play. The champion takes 30 minutes of thinking (see the post on picking the right first champion if you have not yet). The task picking is 20 minutes if you apply the hated-task rule. The folder build is the 90 minutes I have just described.
That is roughly two and a half hours of leadership time, total, before day one of the play. Most of the rollouts I have seen stall did not skip the licenses or the training schedule. They skipped the 90 minutes.
If you would like the worked version of what the starter folder looks like in practice — the role description, the voice file, the task description, the sample inputs and outputs, all from a wine club concierge — that is the next post in this series. It is the most copy-able piece in the whole 20-post run. Coming next: the wine club concierge's AI starter folder, line by line.
If you would like to walk through this on a working call with me, the webinar on Employee AI Engagement for Wineries covers the starter folder build, the hated-task rule, the 3x-a-week habit threshold, and the 60-day play that holds them all together. It is built for winery CEOs, GMs, and DTC directors who are about to roll out AI to staff and want to do it once.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.