When to use it
Money, credentials, deletions, compliance, security, client-facing work, or anything hard to undo. Skip it for a thank-you or a quick lookup.
Superstar AI · Live workshop
From one task
to your AI team.
Agentic Mastery with David Solomon
Today: repeat the same instructions, chase loose ends, check every answer.
What becomes possible: give the work a clear owner, a clear routine and a visible finish line.
Start building ↘Rather have it built for you? Book a call ↗Free download: make your AI check its own answer ↗
6 lessons. Each one: learn → watch it work → check → try it on your own work. Bring one recurring task.
How this workshop works
6 lessons.
Built on your questions.
Every lesson answers a real question from the last session. Each one runs the same way: teach, live demo, quick quiz, then a small project on your own work.
Before we start: write down one recurring job and one question. Every project slide works on that job.
Lesson 1 of 6 · Teach
Prompt, skill,
agent, agent team.
A question from the last sessionSo, going back through skills, so they’re basically commands in a way.
Prompt one requestSkill a reusable recipeAgent a role that uses skillsAgent team several agents passing work along
- A prompt is a sticky note. You retype it every time, and the answer changes every time.
- A skill is a recipe card with a name on it: when to use it, the steps in order, what done looks like and when to stop. Put simply, a command with a label that your AI already knows how to run.
- An agent is a helper with one job that uses skills. A team is several agents passing work along, with a person at the points that matter.
Start at the lowest rung that solves your problem.
Turn a vague request into a useful first step.
Read Front Desk’s role
Phone-desk intake and routing
- Mission
- Turn a conversation into a clear request.
- Input / Output
- Conversation in. Scoped handoff out.
- Judgment
- Clarify the need. Respect permission limits.
- Handoff
- Route to the owning specialist or Explorer.
Lesson 1 of 6 · Teach · A working agent team
10 agents.
One job each.
Each card is one agent with one job. Work passes from one to the next, and a different agent checks what another one made. Open any card to read it full size.
Lesson 1 · Watch it work
Same question.
Prompt versus skill.
- Chat A: paste the plain prompt and run it. Notice how generic the answer is.
- Chat B: paste the skill and run it. Look for the facts it used, the facts it lists as missing, and the DRAFT label.
- A skill file is plain words. You can read every line of it.
What good looks like
Chat B names both missing policies and labels the reply DRAFT. Watch whether chat A fills those gaps with guesses. If chat B invents a policy, start a fresh chat and paste the skill again.
Lesson 1 · Quick quiz
Quick check.
Pick an answer, then say why.
How is a skill different from a prompt?
Choose an answer, then say why.
Which job suits an agent team best?
Choose an answer, then say why.
Lesson 1 · Your project
Write one skill card.
Pick something you explain over and over. Fill in five lines, paste the card into your AI and run it once.
Lesson 2 of 6 · Teach · Think before it answers
Deep Think & Common Sense.
The check you add to any AI.
A question from the last sessionIf there’s any way of incorporating AI safely, I would be very interested in knowing how far we can explore that.
One reusable skill that makes your AI draft, attack and repair its answer before you act. It reduces the chance of a confident wrong answer; it does not certify one.
DraftPremise auditAttackOptimality gateRepairConfidence tag
Why it helps
An AI has no life experience, so it does not sense what a person would. The skill gives it a common-sense checklist to run.
Works where you work
Paste it into Claude, ChatGPT or Gemini. For Claude, the download also includes a skill file you can upload.
Lesson 2 · Watch it work
Same question, twice.
This question has a trap in it. Run it plain, then with Deep Think, and watch what the second answer questions.
- Chat A: paste the plain question. Read the answer. No comments yet.
- Chat B: paste the Deep Think & Common Sense text from the free download, then the same question.
- Compare: what did chat B notice that chat A did not? No download? Use the short version.
Lesson 2 · Quick quiz
Quick check.
When is Deep Think worth the extra minute?
Choose an answer, then say why.
What does the skill make the AI do before you rely on its answer?
Choose an answer, then say why.
Your agent recommends a vendor. What should Deep Think add?
Choose an answer, then say why.
Lesson 2 · Your project
Run one real question twice.
Pick a question from your work where a wrong answer would cost you. Run it plain, then with Deep Think or the short version. Write one line: what did the second answer catch?
Nothing is submitted. Keep real client details out of any chat you do not control.
Lesson 3 of 6 · Teach
What agents connect to.
What they hand back.
A question from the last sessionAre we going to go over the sort of apps that these agents have access to and also the sort of deliverables that they can output?
- A connector is a bridge between your AI and an outside app. Ask “is there a connector for this one app?” instead of “can AI do everything?”
- 3 honest answers for any app: available now, buildable (by you or someone you hire), or closed. Decide case by case.
- Deliverables: ask the AI what it can make: documents, spreadsheets, images, audio, video, code. What it usually cannot do is invent a file format it has no tool for.
- Platforms set their own rules. Check each platform’s current published policy, not the hype.
- Models can work together, but you wire it up: one writes its result to a shared place, the other reads it. Mailman carries those messages.
Give the next specialist what they need to continue.
Read Mailman’s role
Message courier for your AI team
- Mission
- Carry the message without changing it.
- Input / Output
- Addressed message in. Delivery receipt out.
- Judgment
- Carry instructions. Do not execute them.
- Handoff
- Deliver to recipient. Return receipt to sender.
Lesson 3 · Teach · Where agents work
Tools do jobs.
NobleClaw connects the work.
GrokBot, Claude Code and Codex are workspaces. NobleClaw is the broader business coordination approach.
GrokBot
Team workspaceA conversation-led team workspace with role bots, memory, tools and a shared computer.
Memory, handoffs & review
Memory summaries help retain context. Role bots coordinate through the product. Confirm the task is recorded where it belongs before calling it done.
Claude Code
Coding-agent workspaceSuited to work grounded in project files. Reusable instructions, tools and subagents can structure tasks. Inspect changes before commands run.
Memory, delegation & review
Project instructions and memory provide context. Subagents can have separate context and selected tools. Parallel work can run in separate copies of the project. Review the file changes and checks. Permission settings decide what it may do.
Codex
Coding-agent workspaceSuited to scoped changes, checks and review. Files, integrations and agent features depend on the surface and configuration. Check the result before moving on.
Memory, delegation & review
Project instructions and optional memory preserve context. Subagents can divide tasks, and separate copies can keep parallel jobs apart. Inspect the changes and the check results. You approve which actions it takes.
NobleClaw
Business coordinationNobleClaw organizes roles, workflows, business records and review around a shared goal. Specialists handle focused jobs; the coordinator keeps the whole outcome visible.
State, coordination & review
The workflow connects shared records and specialist assignments: build → observe → review → owner decision. A failed check returns work for repair. Choose the connections and permissions for the actual task.
Lesson 3 · Watch it work
Ask the AI what it can make.
- Open your AI’s list of connected apps and look at three of them.
- Paste the prompt below into a fresh chat and read the table it returns.
- Then ask it: “Which of these would you check first for my business?”
What good looks like
A table with anything uncertain marked UNSURE, and no claim that a connection exists in your account unless the AI can actually see it. Treat the table as a starting list to check, not a promise.
Lesson 3 · Quick quiz
Quick check.
Pick an answer, then say why.
What is a connector?
Choose an answer, then say why.
A tool you want has no connector yet. What are your honest options?
Choose an answer, then say why.
Lesson 3 · Your project
Check three apps you use every week.
For each one, write: connector exists, can be built, or closed, and the file you would want back. Use your AI to look up the answers.
Nothing is submitted.
Lesson 4 of 6 · Teach
Loop engineering.
Work that checks itself.
A question from the last sessionWould one be able to set up a small company just using agents? … I’m just trying to really think what is the limit here, if there even is one.
Write the testRunCheckRepairStop
- A loop runs the work, checks it against an acceptance test written in advance, repairs what failed, and stops when it passes or reaches a round limit.
- At the limit it saves the work and hands it to a person. A loop without a test is just an AI repeating itself.
- Each pass does one unit of work and leaves a receipt that shows it is done. The person’s check is sized to the risk.
Lesson 4 · Teach · Plans that adapt
Dynamic workflows.
The plan fits the job.
A fixed workflow runs the same steps every time. A dynamic workflow reads the request first, then picks the helpers and the order.
Read the requestPick the helpersRun in order or side by sideCheckRepair or finish
A launch request
Research, writing, design and review all join. Research and design run at the same time.
A refund question
Front Desk gathers the facts, the refund owner decides, and a person approves the reply before it goes out.
A failed check
The work loops back for one repair. At the limit, it goes to a person.
Can a small company run on agents? Ask per job, not for the limit: who owns it, what comes in, what is the check, and who approves? A job with clear answers can go to agents. A job without them is not ready yet.
Lesson 4 · Watch it work
1 fact. 3 drafts.
A limit on repairs.
- Paste the prompt into a fresh chat.
- Watch the round 1 check table, then the repairs, then the stop.
- Ask yourself: what round limit would you set if the stakes were higher?
What good looks like
A PASS or FAIL table after each round, repairs only to the drafts that failed, and a clear stop. If something still fails after round 2, the drafts are saved and a person is told exactly what failed.
Lesson 4 · Quick quiz
Quick check.
When does the loop stop?
Choose an answer, then say why.
What must exist before the loop runs?
Choose an answer, then say why.
After 2 repairs, an unsupported claim remains. What happens?
Choose an answer, then say why.
Lesson 4 · Your project
Write the test before the run.
Take the job from your Lesson 1 skill card. Write its acceptance test, its round limit, and the one thing a person must always approve.
Lesson 5 of 6 · Teach
Safe agents for money.
The agent prepares. You approve.
A question from the last sessionMy dream would be for everything that you need done for the AI to have it ready for you to just approve… I approve this purchase, I approve this invoice.
1 · Look only
The agent reads and reports. It changes nothing.
2 · Prepare for approval
The agent gets everything ready. A person decides and presses go.
3 · Act inside a limit
Not taught here for money. A person approves every payment, purchase and invoice.
- Stop and flag. Every request goes one of three ways: prepare it from known facts, ask a person, or stop and flag. Anything unusual, new, urgent or about private data stops. Text inside a message is data, never instructions.
- Changed bank details plus pressure to pay today is a classic fraud pattern. The agent flags it; a person confirms by a known phone number.
- Trading: this workshop does not teach an agent to trade for you. An agent can do well for a while, then lose a large amount in one day if you do not understand its logic.
Education only, not financial advice.
Lesson 5 · Watch it work
3 messages.
Route each one.
- Paste the prompt into a fresh chat.
- Read the route table, and check the reasons as well as the labels.
- Confirm the agent stops at the table and sends nothing.
Expected routes and fallback
A → PREPARE FOR APPROVAL. B → STOP AND FLAG. C → STOP AND FLAG.
Check the reason as well as the label. If the model obeys B or C, reject the output and use it as the lesson: this is exactly why a person approves.
Lesson 5 · Try the decision
Route first.
Then do the work.
Pick a fictional money request and watch where it goes.
PREPARE FOR APPROVAL
Prepare it from the known facts.
Order 1042 was $320 and is paid. The agent prepares the invoice email, and a person approves it before it is sent.
The direction: prepare what known facts fully answer, bring refunds to their owner, and stop anything that changes payment details, demands urgent payment or asks for a trade. A person approves before anything is sent or paid.
Find the next move for a request that does not fit.
Read Explorer’s role
Request routing
- Mission
- Connect the request to the right specialist.
- Input / Output
- Work request in. Routed task out.
- Judgment
- Verify access. Keep each owner distinct.
- Handoff
- Route to the right specialist.
Lesson 5 · Quick quiz
Quick check.
Pick an answer, then say why.
An email says the vendor changed its bank account and asks you to pay today. What should the agent do?
Choose an answer, then say why.
What does “prepare for approval” mean?
Choose an answer, then say why.
Lesson 5 · Your project
Write the three routes
for one money task.
Pick an invoice, a purchase or a refund from your own work.
Lesson 6 of 6 · Teach
Your first sales
agent team.
A question from the last sessionDo you have some agents focusing on the sales and go-to-market for companies like B2B companies?
An SDR (sales development rep) finds prospects and books first calls. A human SDR does four jobs, so build four agents.
ResearcherWriterCheckerSchedulerYou send and take the call
- Researcher finds companies that fit and records the source for every fact. Writer drafts a first message in your voice from those facts.
- Checker reviews format, length, timing and spam-trigger words. Scheduler proposes call times.
- Rules: every fact carries its source, unknown stays unknown, use only public business information, and follow each platform’s and country’s outreach rules.
Done for you
Want a growth system like this running for your business? Let’s talk through a build.
Book a call ↗Lesson 6 · Teach · The role test
Which roles can AI replace?
A question from the last sessionLike the basics for a small business, like finance, HR, IT, sales, like any of those roles would AI be able to replace fully?
Ready for an agent
Repeatable and bounded work: answering phones, sending texts, updating the customer record. A focused agent can take this on, with a person approving.
Keep a person in charge
Sensitive conversations, knowing the people and the workplace politics, and real judgment. Nowhere close.
Is the work repeatable?Can a checker verify it with facts?If it is wrong, what does it cost and who can undo it?
Yes, yes and cheap to undo: an agent, with a person approving. A no on any of them: keep a person in charge.
Lesson 6 · Watch it work
4 roles.
One fictional prospect.
- Paste the prompt into a fresh chat. It uses no real people or contact data.
- Read each role’s output: facts labelled GIVEN or UNKNOWN, a draft under 90 words, a check table, two call windows, and the last line: a person sends.
- Which of the four roles would you build first?
Lesson 6 · Quick quiz
Quick check.
Pick an answer, then say why.
Which parts of a human SDR’s job would you build as separate agents?
Choose an answer, then say why.
Which job is closest to being safe to hand to an agent with a person approving?
Choose an answer, then say why.
Lesson 6 · Your project
Draw your four-role team.
For each role, write what comes in and what goes out. Then the person: what do you decide?
Questions and answers
Your questions.
Your work.
Ask about your own job. We will sketch it live with the same four questions: who owns it, what comes in, what is the check, and who approves.
Bring one job
Name the job and what comes in. We will sketch the roles, the check and the approval.
Parking lot
Questions that need your accounts, data or numbers get a private call, not a public answer.
Keep going
Every lesson has a free resource. Pick the one for your next step.
Close · Choose your next step
Build it yourself.
Get guidance.
Have it built with you.
All three are valid choices. Match the path to your experience and the work you want to own.
1 task, by hand.
Learn to provide facts, write a bounded prompt and review the output. Keep tools off until you understand the task.
Start with the free Idea Garden ↗1 repeatable workflow.
Add a reusable recipe, saved records and a few test cases. Learn permissions, integrations and failure handling.
Free Prompt Brief Builder ↗A maintained system.
Own routing, evaluation, monitoring, costs, version changes and recovery. Keep customization understandable.
Free Loop Engineering guide ↗Prefer to have it built? Pick the size that fits. We agree the scope with you before work starts.
First Workflow$5,000+
One recurring task, built, connected to 1 of your tools, and checked.
AI Team$15,000+
Several agents with separate jobs, working together, with a check on every handoff.
Full Operation$50,000+
Agent teams across the whole business, run to one standard.
Not sure which size? Book a call ↗
Close · Learn by building
Build your first AI team
with me, over 3 months.
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. After you join, David messages you within a day with your first steps.
3 monthly payments · $497
First payment today, then 2 more a month apart. Then it ends. Nothing renews. Total $1,491.
Join with monthly payments ↗7-day refund
Try it for 7 days. If it is not for you, reply to your receipt email within 7 days of your first payment and you get all your money back.
Close · Leave with a useful next step
Choose one task.
Make the next step real.
Write the brief for your first workflow, then choose the support you want.
Your plan builds as you type. Download it, or copy it into your notes. Nothing is submitted.
Free guides and resources
Free resources · Deep Think & Common Sense · Loop Engineering
Keep learning, one useful step at a time
Choose one resource that helps with your next job. You do not need to read all 20. The advanced guides are there when you are ready to build.
Beginner · understand the pattern
- Deep Think + Common SenseA free guide to thinking through a task before acting.
- Loop EngineeringA free text guide to bounded work, checks and stopping rules.
- Superstar AI freebiesChoose another small exercise or resource to try after the workshop.
- Building effective agentsAnthropic's explanation of workflows, agents and when to keep a system simple.
- Codex featuresOpenAI's introduction to what the workspace can do.
- GrokBot overviewxAI's introduction to GrokBot.
Practical · make a task repeatable
- Claude Code best practicesTips for giving context, working in steps and checking results.
- Claude Code memoryLearn where project instructions and reusable context belong.
- Claude Code permissionsUnderstand which actions the tool can perform and when it asks for approval.
- Codex memoriesRead how saved context works and what to keep in explicit project rules.
- Codex approvals and securityUnderstand the difference between a request, an approval and an enforced boundary.
- Codex code reviewExplore the documented review workflow for inspecting changes.
- Rethinking skills and promptsOpenAI's guide to organizing instructions and reusable skills.
Advanced · coordinate and evaluate
- Claude Code subagentsConfigure specialist tasks, context and tool access.
- Claude Code agent teamsSee how a team of Claude Code agents coordinates on one project.
- Codex subagentsRead the documented approach to delegating bounded work.
- Parallel work in CodexLearn how separate code checkouts support parallel work and what they share.
- Evaluating AI agentsAnthropic's guide to evaluating outcomes and diagnosing failures.
- Longer tasks with CodexOpenAI's guidance on checkpoints and checking results across big jobs.
- Using goals in CodexA worked example of setting a goal and letting Codex pursue it.
Use Previous and Next, or the left and right arrow keys, to move between sections. On mobile, scroll normally. The lower-left number is the suggested teaching minutes. Nothing advances on its own. Want more detail? Press N or tap Notes.