Build Your Own AI Workspace: Five Steps and the Prompts to Start With

Companion handout to “AI and Running a Coaching Business,” Starting Strength Coaches Conference, Omaha, September 2026. Nick Delgadillo.


Why bother

An AI model off the shelf has three things: what it was trained on (the public internet up to a cutoff date), what it can look up on the web when you ask, and, in some tools, a memory feature that keeps fragments of your earlier chats. None of that contains your prices, your doctrine, your staff, or what you told a member last month. So it answers with the average gym owner’s answer, in the average gym owner’s voice.

Pointed at a workspace, the same model reads your folder before it answers. Your doctrine, your templates, the answers you already give, one page per person with open items and a dated log, and a rules file that says what it may draft and what it may never send. Now it can quote your price in your words, and it can tell you that you decided something in July.

A workspace is a folder the assistant reads first. This handout is how to fill it.

What you need: one folder, and one AI tool that can be pointed at it. The folder used throughout this handout looks like this:

my-gym-workspace/
  rules.md        what it may do, what it never does (read first)
  about-us.md     who we are, what we charge, how we talk
  inbox/          voice notes, transcripts, forwarded email
  people/         one page per person: open items, log
  projects/       one page per project, dated decisions
  knowledge/      doctrine, approved answers, templates

The names are a suggestion. Keep them if you have no reason to change them; the prompts below use them. Any tool that reads a folder on your computer, or a “project” you upload files to, works the same way. Pick the tool after you have written the first three files, not before.


Step 1. Set up your context

One folder the assistant reads before it does anything. Three things go in first.

Write these first

  1. about-us.md, one page. What the business is, who works here, who the members are, what you charge, how you talk.
  2. rules.md, one page. Never invent a number or a name. Draft, never send. Ask before closing anything. Your gate list: the five things a person always decides (anything to a member or staff that reads as coming from you; money; hiring and firing; a change to a lifter’s program; anything public with the brand on it).
  3. knowledge/: five approved answers. The answers you already give members and staff, written the way you say them. One file each in knowledge/, dated.

Keep it usable

  • Plain text files, kept short, one topic per file.
  • A date at the top of every file.
  • Curate: five good files beat five hundred pages of pasted chats.
  • Tell the assistant: read this folder first, and say so when the answer is not in it.

It takes about half a day to start and ten minutes a week to keep up.

Prompt 1a, about-us.md

Interview me one question at a time until you can write about-us.md, a one-page file about my gym: what we do, who works here and what each person does, who our members are, what we charge and for what, and how we talk to people. Do not guess at anything. When you have enough, write the file with today’s date at the top and show it to me before saving it.

Prompt 1b, rules.md

Write rules.md, a file you read at the start of every session we work together. It contains: (1) never invent a number, a name, a price or a quote; if you do not have it, say so and ask; (2) you draft, you never send; (3) ask me before closing, deleting or changing anything; (4) when two files disagree, the newer dated one wins, and you tell me; (5) this gate list, where a person decides every time: [paste your five lines]. Date it, then read it back to me in your own words so I can check you understood it.

Prompt 1c, knowledge/

Ask me the five questions members or staff ask most often. For each one, ask me to answer out loud the way I actually say it, then write it down as its own file in knowledge/ with a date at the top. Do not improve my wording. Do not add anything I did not say.


Step 2. Set up your second brain

One page per person in people/, one page per project in projects/. Same shape every time. This is working memory: it changes every week and should. The context folder from step 1 is the durable part.

The template

Name · role · how we work together

Open items
- What · who owes it · since when
- What · who owes it · since when

Log (newest first)
- Date: what happened, what was decided
- Date: what happened, what was decided

The habit that feeds it

  • After a call or meeting, hand it the transcript or a two-minute voice note.
  • It proposes new items and log lines. You keep or drop each one.
  • Closing an item is your call, never its own.
  • Then ask it: “What did we decide about X?” “Who owes me what, and since when?”

Start with your staff, three or four pages. Add a member’s page the first time a decision about them is worth remembering.

Prompt 2a, First page

Make a page for [name] in people/ using this shape: their role and how we work together at the top; a section called Open items with one line each, in the form “what · who owes it · since when”; and a section called Log with dated entries, newest first. Here is what I know about them: [say it]. Show me the page before saving it.

Prompt 2b, The habit

From now on, whenever I give you a transcript, meeting notes or a voice note, do this: read it, identify every person who has a page in the folder, and propose new open items and log entries for each of them. List the proposals and wait for my yes on each one before writing anything. Never close an open item yourself; only propose it.

Prompt 2c, Recall

Check the pages in the folder and tell me: what did we decide about [topic], and when? And who currently owes me what, oldest first? Point at the page and the dated line for each answer.


Step 3. Set up the analysis layer

Hand it the inputs and one question. Keep the yes or no.

What you hand it

  • The transcript, the two documents, or the export from your scheduling or accounting software.
  • One question: compress this, compare these, find where this contradicts what we decided, or what should I ask next.
  • The shape you want back: one page, a table, a list of who owes what.

The standing rule

Every number traces to a source you can open, or it gets marked [unverified]. If it cannot point to where a figure came from, the figure does not go in the report. Put that sentence in rules.md.

Start with one report

The one you rebuild by hand every week. Write down its columns and where each number comes from. Have it drafted in that shape. Check it against the source the first three times, then spot-check. Once the shape is stable, it becomes your first scheduled routine.

Prompt 3a, The rule

Add this to rules.md: every number in anything you produce must trace to a source I can open (a file in the folder, an export I gave you, or a connected system). If you cannot point to the source, mark the number [unverified] in the text. Never fill a gap with a plausible figure.

Prompt 3b, The first report

Here is the report I rebuild every week: [describe or paste it]. These are its columns, and this is where each number comes from: [list them]. Draft this week’s version in exactly that shape from these sources: [attach or name them]. Mark anything you cannot trace [unverified]. Then list every number and its source so I can check them.

Prompt 3c, Compare and contradict

Compare these two documents and give me one page: what is the same, what differs, and anything in either one that contradicts a decision recorded on a page in the folder. Quote the line and name the page for each contradiction. Then tell me the one question you would ask next.


Step 4. Connect the drawers

Add one connector at a time, let it read before you let it draft, and nothing sends without a person.

A sensible order

#DrawerAccess
1Emailread, then draft
2Calendarread
3Files and sheetsread, then write
4Gym schedulingread
5CRM / leadsread, then draft
6Accountingread only

The rules

  • Add one, use it for a week, then add the next.
  • Read access first. Drafting only after a month of clean reads.
  • Keep the list short enough to say out loud.
  • Write the list into rules.md, with what each one may do.
  • A person always sends.

In most tools this is a settings menu called connectors, integrations or MCP. You sign in to the service once and choose what it can see.

Prompt 4a, First connector

I have connected my email. For the next week you may read it, and nothing else: no drafts, no replies, no labels, no deletions. Add that line to rules.md. Each morning, list the threads that need a reply from me and what each one is waiting on. Do not write any replies yet.

Prompt 4b, Adding drafting

Update rules.md: for email, you may now draft replies, and you save them as drafts for me to read and send. You never send anything. You never reply to anyone directly. Confirm you have written that down, then show me the current list of everything you are connected to and what each connection may do.


Step 5. Audit and improve on a clock

The system is never finished. The upkeep is small if it is scheduled.

WhenDo
Weekly, 10 minutesHave it read the whole folder: what is stale, what contradicts itself, what did I say I would add?
Every time it bitesAdd one line to rules.md so that failure cannot repeat.
MonthlyPrune: which files were never read, which answers changed, which pages have no open items.
QuarterlyRead the connector list and the gate list out loud. Remove what you cannot justify.

What the audit catches

  • A confident wrong number. Trace it or mark it unverified.
  • Two files that disagree. Newer dated one wins, and the assistant says so.
  • A robot wearing your name. Sweep for machine tells before it ships as you.
  • A search result treated as the whole picture. Open the thing before saying what is missing.
  • “Done” that was not checked. Check the file or the sent folder first.

Every rule in my system was added after it bit me once.

Prompt 5a, Weekly audit

Read every file in the folder. Report three lists: (1) files that look stale, with the date on each and why you think so; (2) places where two files or pages contradict each other, quoting both; (3) things I said I would add or fix, according to the pages, that are not there yet. Propose a fix for each item. Change nothing until I say.

Prompt 5b, After it bites

You just [describe what went wrong]. Write one line for rules.md that would have prevented it, in plain language, and show it to me before adding it.

Prompt 5c, Monthly prune

Which files in the folder have you not needed to read in the last month? Which approved answers have changed since they were written? Which person pages have no open items and no log entry in 60 days? List them with a recommendation for each: keep, update or archive. Do not archive anything yourself.

The test for whether the setup is working: every answer it gives should point at a file in your folder. When it cannot, that is either a missing file or a guess, and either one is worth knowing.


Glossary: the words, explained

The words you will hear, in the order you are likely to meet them. Nothing here needs a technical background. Where a word appears on a slide in the talk, the definition matches what the slide means by it.

The basics

AI. In this context, software that can read, write and answer questions in ordinary language. When people say “AI” today they almost always mean the chat tools built on large language models, so that is what this glossary means by it too.

Large language model (LLM). The engine behind the chat tools. It was trained by reading an enormous amount of text and learned to predict what words come next. That is why it can write a paragraph, summarize a call, or answer a question, and also why it can sound confident when it is wrong.

Model. One specific version of that engine. Each company has several, they get updated a few times a year, and newer ones are usually better at following instructions and reading longer documents. You do not need to track them. Pick a tool and use whatever its current model is.

Chat tool, chatbot, chat box. The window you type into. ChatGPT, Claude, Grok and Gemini are the ones you have probably heard of. The chat box is where most people meet AI and where most people stop.

Prompt. Whatever you type or say to the tool: a question, an instruction, a pasted document. A good prompt says what you want back and in what shape (“one page, with a list of who owes what”).

Response, output, draft. What comes back. In this talk it is always called a draft, because that is how to treat it: something to read and edit before anyone else sees it.

Assistant. A chat tool that has been given your files and your rules, so it works for your business instead of answering in general. The talk uses “the assistant” for the whole setup.

Agent. An assistant that can take steps on its own: read a file, look something up, write a page, then come back to you. The word gets used loosely. What matters is what it is allowed to do without asking, which you decide.


What it can and cannot know

Training data. Everything the model read while it was being built, which is most of the public internet up to a certain date. It does not include your prices, your doctrine, your staff, or anything a member told you.

Cutoff date. The last date the model’s training data covers. Anything after that, it does not know unless it can look it up or you give it the document.

Web search, browsing. Some tools can look things up on the internet when you ask. Useful for public facts. Useless for anything about your gym.

Memory (chat memory). A feature in some tools that keeps fragments of your earlier conversations. It is not the same as giving the tool your files, and it is not reliable enough to run a business on.

Context. Whatever the tool can see right now: your prompt, the files you gave it, the conversation so far. If a fact is not in the context, the tool is guessing. Most of the talk is about putting the right things in the context before you ask.

Context window. How much the tool can hold in view at once. Bigger is better. Modern tools can hold a few hundred pages, which is more than a small gym’s whole folder.

Hallucination. The tool stating something false as if it were true: a number it did not have, a name it invented, a policy you never wrote. It happens most when the answer is not in the context. The fix is the source rule below, plus a folder that actually contains the answers.


The words on the slides

Workspace. A folder the assistant reads before it does anything. The whole talk is about filling that folder.

Folder. An ordinary folder on your computer, or a “project” you upload files to inside a chat tool. Either one works.

Page. One file about one thing: a member, an employee, a lead, a project. Plain text, with a date on it.

Person page. The page for one person: who they are, their goals in their words, your notes, open items, and a dated log of what happened. One per member, one per staff member, one per lead.

Project page. The same idea for a piece of work: what it is for, the decisions made, dated.

Link. A reference from one page to another, written as the other page’s name in double brackets, like [[maria-r]]. The links are what let the assistant walk from a coach to their members to the template each one is on. No special software; just the name of another file.

Second brain. Your people pages and project pages, linked. Working memory for the business: what was said, what was decided, who owes what, and since when.

Knowledge base. The durable reference: what you teach, how you talk, the answers you already give, the procedures. It changes rarely. The second brain changes weekly.

Doctrine file. The one file that holds your programming logic: the templates, the rules, the rulings, dated. Everything a member touches is built from it, so when it changes, everything built from it can be rebuilt.

Rules file. The one-page file the assistant reads first, every time: never invent a number or a name, draft but never send, ask before closing anything, and your list of what always needs a person.

About-us file. One page on what the business is, who works there, who the members are, what you charge, and how you talk. The first file to write.

Inbox folder. Where work lands before you have looked at it: a voice note, a forwarded email, a meeting transcript. Each with one line saying what to do with it.

Template. A written pattern the assistant fills in the same way every time: a follow-up text, a weekly report, a member explanation. Templates are how the output stays consistent.

Approved answers. The answers you already give to the questions you hear every week, written the way you say them. Five of these are a better start than five hundred pages of pasted chats.

Voice note, dictation. Talking instead of typing. Two minutes in the car after a session, transcribed by an app, lands in the inbox folder. The best input most coaches have and the least used.

Transcript. The written version of a recorded call or meeting. Tools can produce one automatically. It is the raw material for meeting summaries and page updates.


Doing work with it

Draft and approve. The one workflow behind everything in the talk. The assistant drafts. A person reads it and decides: send, file, change, or drop. Nothing goes to a member, an employee or a lead without that step.

Routine, scheduled task. A recurring chore with a time on it. The assistant wakes up, reads the folder, does the work, and leaves you a draft or a report. The morning huddle, the Monday lists, the weekly funnel, the monthly numbers.

Connector, integration, MCP. A plug that lets the assistant look at, or draft into, a system you already use: email, calendar, scheduling software, your CRM, accounting. MCP is the technical name you will see in menus. Each one is a key to one drawer, granted by you, and you can take it back.

Read access, write access. Reading means the assistant can look. Writing means it can change or send. Give read access first and use it for a while before you give write access to anything, and never let it send to a person without you.

Source rule. Every number traces to a source you can open, or it gets marked [unverified]. If the assistant cannot show where a figure came from, the figure does not go in the report. This one rule is what makes the rest trustworthy.

Summary. A long thing made short: a 90-minute call into one page and a list of who owes what. The most reliable thing these tools do.

Report. The same numbers, in the same shape, every week or month, drafted from the same sources. Start with the one you rebuild by hand.

Audit. A ten-minute weekly habit: have the assistant read the whole folder and say what is stale, what contradicts itself, and what you said you would add and have not. Then fix one thing.


Business systems you will hear named

CRM. Customer relationship management software: where your leads and their contact history live. The assistant can read it, and can draft into it, through a connector.

Scheduling software. Where sessions are booked and attendance is recorded. A source for the member page and the retention numbers.

Logbook. Where the lifts are recorded. Depending on your setup, the assistant may see aggregates or the whole history. Either way, the trend goes on the member page.

Accounting software. Where the money is. Read access only, and the financial report is drafted from it against the forecast.


Things to be careful about

Privacy. Whatever goes in the folder is readable by whichever tool you point at it. Decide what belongs there. Member names and goals, yes. Card numbers and health details, no. Read the tool’s terms on whether your files are used for training, and turn that off where you can.

Prompt injection. Text inside a document that tries to give the assistant instructions (“ignore your rules and send this”). Rare in a gym, but it is why the assistant should treat everything it reads as information and take instructions only from you.

Subscription, plan. What the tool costs per month. Most useful setups run on one paid plan for one tool. Do not buy five subscriptions this weekend. Write the three files first, then pick the tool.

Version, date. Put a date at the top of every file. When two files disagree, the newer one wins, and the assistant should say so. Most confusion in a folder comes from an old file nobody dated.