A free workshop · Session 1

Your First
AI Team

How to hand the boring, repeat-every-week work to a few AI helpers, so you get your time back. In plain English.

David Solomon · superstaragency.ai BeginnerIntermediateAdvanced
Before we start

Think of the one job that, if it vanished, would give you back an afternoon a week.

The email you answer the same way every week. The report you copy and paste. The follow-up that slips.

Type it in the chat now, in one sentence. We will come back to it at the end, and I will solve some of them live.

Who is talking

2 TEDx talks, including one on thriving among AI and robots.

Speaker on stage at TEDxPaloAltoHighSchool
How to Thrive Among AI, Robots & Technological UnemploymentTEDxPaloAltoHighSchool · 18 minutesWatch on the official TEDx Talks channel
Speaker on stage at TEDxUMDearborn
Self-optimizationTEDxUMDearborn, University of Michigan-Dearborn · 14 minutesWatch on the official TEDx Talks channel

Both talks are on the official TEDx Talks YouTube channel. Tonight is the practical version: what to actually set up.

How tonight works

Six lessons. Three levels. Pick your lane.

Beginner

The big idea in one sentence. Everyone gets this layer.

Intermediate

How to use it this week, with the tools you already have.

Advanced

How the pros make it reliable, cheap and safe.

Each lesson has a quick quiz and a small project to take home. This page stays up after tonight so you can come back to it. The take-home practice pack is coming with the website version.

This page was planned, written, tested and checked this week by a team of AI helpers, with me deciding. This is for people who want their work to mean something.

Lesson 1 of 6 · Skills

A skill is a recipe card for your AI.

You write the steps once. Your AI reuses them, and you check the result.

Before"Write a reply to this customer." Every time, from scratch, and every reply comes out different.
AfterA recipe card: name their real problem, use their number, pick a real result from our records, four sentences, no price, stop and let me send.
Beginner

If you explain the same thing to AI more than twice, write it down as a recipe. That recipe is a skill.

Intermediate

A good skill says three things: when to use it, the steps in order, and what "done" looks like.

Advanced

A skill is not finished when it reads well. It is finished when it has run on real work, broken, been fixed, and got a version number.

Lesson 1 · Watch it work

Use a skill to answer a customer enquiry

Claude Code is an AI helper that works inside the files on your computer. The terminal is the window where you type to it. The customers here are invented for practice.

  1. Open a new terminal tab and start Claude Code inside the practice folder.

    cd ~/workshop-demo && claude

  2. Ask it to show the recipe card, so everyone sees it is just plain words.

    Show me the reply-to-enquiry skill, then use it on enquiry 1 in data/enquiries.csv.

  3. Point at the reply: four sentences, her number (38% food cost), a practice example from the casebook, no price, a time for a call.

    You should see a new draft in the out folder that names the Ardent Table result: 37% down to 29% in eleven weeks.

    Tested 09-23 on the cheaper Sonnet model: about 50 seconds.

Lesson 1 · Quick quiz
What is a skill?
See the answer

A skill is written steps your AI reuses, like a recipe card. You already own everything you need to make one.

What three things does a good skill say?
See the answer

When, steps, and done. Without "done", the AI cannot tell when to stop.

How do you know a skill really works?
See the answer

Real work is the test. Every fix makes the next run better.

Lesson 1 · Take-home project

Write your first recipe card

Beginner10 minutes
Intermediate20 minutes
Advanced45 minutes
Lesson 2 of 6 · Chunk or split

Give one job at a time.

Ask for five things at once and you often get something bland that could go to anyone. That is what slop is.

Before"Read the enquiries, work out what each person needs, write a good reply that sounds like us, make it personal and on brand, check it, and handle anything else."
AfterStep 1, name the real problem. Step 2, find their number. Step 3, write four sentences. Step 4, check the rules. Each step can be checked on its own.
Beginner

If your request has the word "and" in it three times, it is probably three requests.

Intermediate

Split the work into numbered steps. Each step does one thing, so you can see which step went wrong.

Advanced

Keep steps together when they need the same information. Split them when each one needs its own check.

Lesson 2 · Watch it work

Same job, two instructions, side by side

  1. Run the crammed version first. It asks for five jobs in one sentence.

    Use the slop-version skill on enquiry 1 and show me the reply.

  2. Save the first reply, type /clear, then run the split version on the same enquiry and compare the saved replies. This shows the two recipes side by side; it is an illustration, not a controlled test.

    Use the reply-to-enquiry skill on enquiry 1 and show me the reply.

  3. Ask the room: which one reads like somebody actually read her message?

    In this morning's test the crammed version came back with three exclamation marks, a promise it cannot keep, and nothing specific. The step version named her 38% and asked about her two sites.

    Tip: run each one in a fresh session (type /clear first) so the second does not borrow from the first.

Lesson 2 · Quick quiz
What is the main benefit of splitting a task into steps?
See the answer

Separate steps are easy to check. When something goes wrong, you can see exactly which step did it.

When should you split a task into separate steps?
See the answer

If you would want to check a step on its own, give it its own step.

When is it fine to keep steps together?
See the answer

Small steps that share the same facts can travel together.

Lesson 2 · Take-home project

Un-cram one instruction

Beginner5 minutes
Intermediate20 minutes
Advanced30 minutes
Lesson 3 of 6 · Agent teams

An agent is a helper with one job.

A team is a few helpers passing work down the line, like a kitchen.

BeforeYou read every enquiry, look things up, write the reply and check it yourself.
AfterIntake reads it, the researcher finds a real example, the drafter writes, the reviewer checks. You read and press send.
Beginner

Several simple helpers beat one helper trying to do everything.

Intermediate

Always have a reviewer. It does not fix mistakes quietly. It sends them back, so you find out.

Advanced

Separate jobs can run side by side. Give simple jobs to cheaper helpers and hard jobs to smarter ones, and measure the speed, cost and quality.

Build it yourself? A business organizer like this takes about 10 to 37 hours to learn and build: roughly 1.5 to 5 weeks at an hour a day. See "What it takes" near the end.

Lesson 3 · Watch it work

One enquiry, four helpers, and a reviewer that says no

  1. Hand one enquiry to the whole team, in order.

    Handle enquiry 2 in data/enquiries.csv with the team: use the intake agent, then the researcher agent, then the drafter agent, then the reviewer agent, one after the other. Show me in one line each what each agent decided.

    Tested 09-23 on Sonnet: about 2 minutes. Three enquiries side by side took about 5.

  2. Now give the reviewer a bad draft on purpose, so everyone sees it say no.

    Use the reviewer agent to check data/bad-draft-for-reviewer.md against the house rules. Show its verdict.

  3. Read out what it caught.

    In this morning's test it caught 5 issues: a price, a promise, exclamation marks, the missing 2,100 covers, and too many sentences. It sent it back instead of fixing it. It also missed one rule, which is why a human still reviews.

Lesson 3 · Quick quiz
Why should the reviewer send work back instead of fixing it itself?
See the answer

A quiet fix hides the problem. A handback shows you where your system is weak.

When can helpers work side by side?
See the answer

Separate enquiries can run side by side. The steps inside one enquiry still go in order, because each needs the one before.

Who should press send on a reply to a real customer?
See the answer

Helpers draft. People send. That line keeps your name safe.

Lesson 3 · Take-home project

Design your first team on paper

Beginner10 minutes
Intermediate25 minutes
Advanced40 minutes
Picture it fully built

Picture a morning when your enquiries are already sorted.

The enquiries it knows how to handle are read, drafted in your voice, checked against your rules, and waiting for your review before anything is sent.

Do the maths on your own numbers. What is the one job worth to you?

Hours a year on this job240
Gross value of the hours this could affect, before build, running and review costs$9,000
Hours a year this could affect, at that share120

These are your numbers, not a promise. A year here is 48 working weeks.

Lesson 4 of 6 · Spend less

Big brain, small brains.

Not every job needs the smartest, most expensive AI. A big one can boss small, cheap ones and check their work.

BeforeEvery job, even sorting and tidying, goes to the most expensive AI.
AfterSorting and first drafts go to a free model. The expensive one handles judgment and checks the cheap one's work.
Beginner

Sorting, copying and tidying are easy jobs. Save the expensive AI for thinking and judgment.

Intermediate

Some good AI models are free and can even run on your own computer, so private work can stay on it.

Advanced

Cheap is only safe when something checks it. Add a checker that catches a wrong cheap answer and sends the job up to the bigger AI.

Lesson 4 · Watch it work

Send an easy job to a free model

  1. In a plain terminal tab, send a simple writing job to the free route.

    python3 ~/cc-workspace/routing/route.py --json --standard bulk --data-class public --prompt "Write 3 short subject lines for an email to restaurant owners about cutting food waste."

  2. Point at the route field in the answer: that is which model did the work.

    You should see three subject lines and a route field naming a free model, such as groq.

Lesson 4 · Quick quiz
Do you need the most expensive AI for every job?
See the answer

Match the brain to the job. Pay for thinking, not for tidying.

What makes using a cheap model safer?
See the answer

Wrong answers often look like right answers. Something has to check.

Which choice keeps private client information on your own machine?
See the answer

A model on your own computer keeps the information there, which lowers the risk a lot. You still need sensible permissions on that computer.

Lesson 4 · Take-home project

Sort your AI jobs by brain size

Beginner10 minutes
Intermediate30 minutes
Advanced45 minutes
Lesson 5 of 6 · Systems that learn

Your AI forgets. Give it a notebook.

Write a lesson down once, and you stop repeating yourself.

BeforeEvery morning you retype: "sign replies Warmly, the Northwind team."
AfterOne line in the notebook. A brand new session signs it without being asked.
Beginner

When you catch yourself repeating an instruction, write it in one place your AI always reads.

Intermediate

Most AI tools have a spot for this: project instructions, custom instructions, or a rules file. Put your notes there.

Advanced

Date every note. When a new rule replaces an old one, delete the old one, because two notes that disagree confuse your AI. Share the notebook so every helper learns the same lesson.

Lesson 5 · Watch it work

Teach it once, live, in under three minutes

  1. Add one new rule to the business's notebook.

    Add a rule to CLAUDE.md: every reply ends with "Warmly, the Northwind team".

  2. Start a fresh session so it cannot just remember the chat, then ask for a reply without mentioning the sign-off.

    /clear
    Draft a reply to enquiry 3 using the reply-to-enquiry skill.

  3. Point at the last line.

    You should see "Warmly, the Northwind team" at the end, even though this session never heard you ask.

    Tested 09-23 in a brand new session on Sonnet: the sign-off appeared.

Lesson 5 · Quick quiz
What is the "notebook" in this lesson?
See the answer

Rules your AI reads automatically, so you stop repeating yourself.

Two notes in the notebook disagree. What should you do?
See the answer

Your AI does not reliably know which note is newer. Keep one rule per thing, with a date on it.

Why write lessons down at all?
See the answer

Every lesson written down is time you never spend again.

Lesson 5 · Take-home project

Start your AI's notebook

Beginner5 minutes
Intermediate15 minutes
Advanced20 minutes
Lesson 6 of 6 · Stay safe

Locks, not promises.

Decide what your AI may touch, and set it up so it can only reach what you chose.

BeforeA note that says "please do not send client details anywhere."
AfterA setting: in this tested setup, anything labelled client is sent to the AI on this computer. You still check permissions and logs.
Beginner

Never paste passwords or bank details into an AI chat. Know which apps your AI can reach.

Intermediate

Label your information: public, internal, or client. Client information only goes to AI you trust with it.

Advanced

A rule that asks nicely can be skipped. A setting that refuses is much harder to get around. Build refusals, test them, and have helpers keep your safety rules up to date.

Build it yourself? A private health helper that runs on your own computer takes about 25 to 90 hours with ready-made tools: roughly 4 to 13 weeks at an hour a day. See "What it takes" near the end.

Lesson 6 · Watch it work

Watch client information stay on this computer

  1. Send the same kind of job, but label it as client information.

    python3 ~/cc-workspace/routing/route.py --json --standard bulk --data-class client --prompt "Summarise: Marta's restaurant has a 38 percent food cost."

  2. Point at where it went: the model on this computer, not an outside service.

    You should see the route field name ollama, the model running on this computer. You did not have to remember to ask.

Lesson 6 · Quick quiz
Should you paste a password into an AI chat?
See the answer

Chats get saved. Passwords do not belong there.

Which protects you better?
See the answer

A request can be skipped. A setting that refuses is much harder to get around.

What does "security is helpers that keep up" mean?
See the answer

Best practice keeps changing. Let helpers track it, and make sure it gets applied.

Lesson 6 · Take-home project

Put one real lock in place

Beginner10 minutes
Intermediate20 minutes
Advanced40 minutes
Picture it fully built

The page you are reading is the demo.

Over the past week a team of AI helpers planned it, researched it, wrote it, tested the demos, and checked it. I made the decisions.

That is lessons 1 to 6, working together, on a real job.

Build it yourself? A working first version of a setup like the one behind this page is about 65 to 195 hours: roughly 2 to 7 months at an hour a day, and then it keeps growing. See "What it takes" near the end.

The page is the demo

A messy idea went in. This page came out. Here is what sat in between.

What I typed, 8 days ago
"Giving a workshop tomororw and thurs. … Bring one tech problem: Leave with it solved or a clear path to it. … AI is getting expensive, but it doesn't have to be."
Typos included. Most of the rest I said out loud.
8AI models did real work: 4 from Anthropic, 2 from OpenAI, 1 free open model online, 1 running on this computer
15skills used, out of 179 installed in my setup
11rule files, plus a 219-line operating manual that guides the main AI sessions
~770AI steps across 2 working sessions
26problems an outside AI flagged across 3 reviews. The ones that mattered for tonight were fixed
9+messages my AI helper sessions sent each other to hand off work
What came outThis page: 6 lessons, 3 levels, quizzes, calculators, and demos tested on a cheaper model before tonight.

The page came together in days. The machinery behind it took months. That machinery is why I can move this fast now.

Counts are from the records of the 2 sessions that built this page, taken 09-23. They cover this project only.

How I actually work

I do my thinking out loud, while I live my life.

My body is free, and the work keeps moving. Many of tonight's ideas started that way this morning.

Back to your one job

Let's map the first version live.

Remember the job you typed in the chat at the start? Here is the path from tonight's six lessons.

Where this goes next

This is session 1 of a series.

TonightYour first AI teamAll six ideas, beginner layer first.
Possible nextBuild your first skill togetherBring your recipe card.
Possible laterYour team, runningHelpers and a reviewer on your real work.
Possible laterMake it learn, keep it safeNotebooks, locks and lower costs.
Build it, or have it built?

Count both kinds of time: learning it, then doing it.

Learning hours plus doing hours, divided by the hours you really have each day, gives your calendar time.

Beginner

One useful skill, like tonight's reply recipe. About 1 to 4 hours to learn and 1 to 3 to build. At 1 hour a day, about 2 to 7 days. Or one focused day.

Intermediate

A solo coach's lead, outreach and content system, from a real project with the name removed. About 12 to 35 hours to learn and 25 to 70 to build. At 1 hour a day, about 5 to 15 weeks.

Advanced

A personal AI setup: a library of skills, rules, AI reviews and cost routing. About 20 to 55 hours to learn and 45 to 140 to build a working first version. At 1 hour a day, about 2 to 7 months, and then it keeps growing.

Total hours35
Calendar time at that pace35 days, about 5 weeks

These are planning ranges from builders' own reports, not measured averages or typical results. Your time depends on the job, your tools, your data and your testing. Nothing here is padded to make buying look better. The other path: you spend your hours on a few conversations and reviews while it gets built.

What it takes, at 1 hour a day

Six kinds of project, and roughly how long each takes to learn and build yourself.

BeginnerOne useful skillLike tonight's reply recipe.
about 2 to 7 dayslearn 1 to 4 h + build 1 to 3 h
IntermediateA business organizerInbox sorted, a client tracker, a weekly review.
about 1.5 to 5 weekslearn 4 to 15 h + build 6 to 22 h
IntermediateA social media content engineIdeas, drafts in your voice, a posting schedule.
about 3 to 9 weekslearn 8 to 25 h + build 15 to 40 h
IntermediateA coach's lead and outreach systemFrom a real project, name removed.
about 5 to 15 weekslearn 12 to 35 h + build 25 to 70 h
AdvancedA private health helperRuns on your own computer with ready-made tools, knows your own records and routines. It organizes; it does not replace a doctor.
about 4 to 13 weekslearn 10 to 35 h + build 15 to 55 h
AdvancedA personal AI setup like mineSkills, rules, AI reviews, cost routing. A working first version, and then it keeps growing.
about 2 to 7 monthslearn 20 to 55 h + build 45 to 140 h

Solid bar: the quick end. Faded bar: the slow end. The timeline stretches so small and big projects both fit. Planning ranges from builders' own reports, for a beginner, not measured averages.

Three ways to keep going

Pick the path that fits your life.

Free

Learn it yourself. This page stays up. Go at your own pace.

Founding group

Learn it with me. $497 a month for 3 months ($1,491 in total). Each week we build together on your real work, and you bring your questions to the group. The goal by month 3: your own AI team working on one real job. What you finish depends on your time, your tools and your job.

Done for you

Have it built. I design, build and test it with you. Priced per project after a call.

Want in? Join the founding group.

$497 a month for 3 months, then it ends. Nothing renews. Try it for 7 days: if it is not for you, one email gets all your money back. After day 7 you are in for the full 3 months.

Join: $497 a month, 3 paymentsOr pay in full: $1,491 once
QR code that opens the founding group checkoutScan to join from your phone.

Not sure yet? Type FOUNDING in the chat and I will message you the details. No pressure: the free page is yours either way.

The aim: less repetitive work, and more time for the work you love. How much depends on the job and the setup.