01 / 15
AI SHERPA  ·  THREDUP AI MAKER WEEK  ·  JUNE 29 2026
AI Sherpa WWW.AISHERPA.ME

1, 2, 3
Claude

The three levels of Claude and what to ask of each one.

Workshop Lead Santiago Restrepo

HOW WE WORK

We put AI to work for people. Not the other way around.

500+ hours of hands-on
AI coaching

We start with you, not the technology. The whole point is to make AI serve the way you already think and work: hands-on, practical, and genuinely fun. You walk out with AI working for you, not one more tool to manage.

Santiago Restrepo Santiago Restrepo AI SHERPA · YOUR GUIDE FOR THE DAY

The room today

This is who's in the room.

From your own registration answers. Almost everyone here already uses AI every day, so we are not starting from zero. We are going further.

Curious, barely tried it
0%
Use it now and then
0%
Daily AI user
78%
Build projects, use Co-Work
22%
Build & automate with Code
0%

Live from registrations as of this morning.

The shape of the day

Two halves. First you master the basics, then we build together.

Morning Master the basics
01Understanding the Claude ecosystem: the models, and the three ways to work (Chat, Cowork, Code)
02Talking to Claude well, and building your master prompt: prompts, brain dumps, Wispr Flow
03Build your first assistant with Projects
04Turn your assistant into a Skill, and build skills from scratch
05A look at what's next: plugins, scheduled tasks, and other advanced features
06Other models worth knowing: Gemini for deep research, ChatGPT for image generation
Afternoon Build together
Building in public: bring a real problem you are working on
We work through it live, screens shared, as a group
Open Q&A: anything you want to put on the table
Watching each other build is how the whole room levels up
01 02 03 04 05 06

Module 01 of 06

Understanding the Claude ecosystem

ANTHROPIC'S REAL PRODUCT

The product is the intelligence itself.

Haiku, Sonnet, Opus, Fable
Haiku Fast and lightweight, for quick high-volume tasks.
Sonnet The balanced default for most everyday work.
Opus The most capable, for the hardest reasoning.
Fable The newest and most experimental, rarely seen.

To let people actually use these models, Anthropic built three environments to work with them.

The Framework

There are four ways to work with Claude.
Most people don't know the difference.

Claude Chat
1 Claude Chat / WEB
Claude Cowork
2 Claude Cowork / DESKTOP
Claude Code
3 Claude Code / TERMINAL
Claude API
4 Claude API
Claude API: a robot answering calls in a call center

4 · API

Claude API

You wire Claude straight into your own software.

An API is how two pieces of software talk to each other, with no person typing in between. Instead of you chatting with Claude, your own tools call Claude directly: to read, write, sort, or generate, at scale and on their own. It is how Claude gets built into a product or a workflow that just runs. We will not go deep today, just know it exists.

Built for: developers and automations

Claude Chat illustration

1 · Web

Claude Chat

You rent a cubicle in their building.

You rent a workspace inside Anthropic's offices. Your password is the key to your account, and your files live on their servers, not yours. Space is limited, but you can reach it from anywhere, even your phone.

Instruction layer: Master Prompt

Chat: instructions, project context, memory, skills, connectors
Claude Cowork illustration

2 · Desktop

Claude Cowork

You bring the mind into your own house.

Instead of going to their offices, you bring Claude home. It works directly in the folders on your computer: no uploads, no size limits. Folder instructions give it a memory that loads every session. Powerful, with guardrails still on.

Instruction layer: Folder instructions

Cowork: global instructions, folder instructions, memory, skills, connectors
Claude Code illustration

3 · Terminal

Claude Code

You hand over the keys to the whole house.

Full access to your machine. It can build, run, and ship, not just talk. The most powerful and the most technical of the three. Memory, skills, and automations all live here.

Instruction layer: CLAUDE.md, skills, hooks

Code: CLAUDE.md hierarchy, memory, skills, connectors

Across Chat, Cowork & Code

What lives where.

Available Partial ○ Not available
CapabilityChatCoworkCode
Standing instructionsa note Claude reads every time
Projects & project filesa saved workspace of context and files
Memory across sessionsit remembers between conversations
Skillsteaches Claude a task, your way
Pluginsa bundle of skills, connectors and commands
Connectors (MCP)a live link to your other apps
Subagents & agent teamsClaude spins up helper Claudes
Hooksauto-rules that fire at set moments
Scheduled tasks (local)recurring jobs on your machine
Cloud Routinesrecurring jobs that run in the cloud
Remote / background agentsstart it, walk away, it finishes
Artifactsa live panel for docs and dashboards
Direct local file accessworks right in your computer's files
Code execution sandboxactually runs code to test things
Computer useclicks, types and navigates a screen
Robot holding a CLAUDE.md note

THE ONE FILE TO REMEMBER

CLAUDE.md: the note Claude reads every time.

Think of CLAUDE.md as a sticky note left inside a folder. Every time Claude opens that folder, it reads the note first: who you are, how you work, what matters. Each surface has its own version of that note.

Chatthe Project instructions
Coworkthe Folder instructions
Codethe CLAUDE.md file itself

Now let's go build with it.

01 02 03 04 05 06

Module 02 of 06

Talking to Claude well, and your master prompt

Demo 1 · Claude Chat

Start at the front door: Claude.ai.

What it is

Claude.ai, in your browser or on your phone. The most accessible way in, and where most people meet Claude.

Where we will take it

Past quick answers. We give Claude a Master Prompt, then build a specialist that does real work.

Like renting a cubicle in Anthropic's building.

Working with Claude.ai in the browser

Demo 1 · The Master Prompt

Tell Claude who you are. Once.

Most people treat Claude like Google: short questions, generic answers. The unlock is context. Answer eleven questions out loud, and Claude writes your Master Prompt.

The 11 questions

Built once, by voice

1Your name and what you do
2Years in your field, and where you have worked
3The two or three projects that define you
4What you are working on right now
5The biggest problem you are solving today
6Your strongest belief about your work
7The tone you want from Claude, and the one you do not
8The role Claude should play: assistant, partner, or coach
9The habits that work against you, and what to flag
10Who you work with and for
11The language you work in
01Answer 11 questions out loud
02Wispr Flow turns your voice into text
03Ask Claude to write your Master Prompt
04Paste it into your profile
A living document. Refresh it after any big change, or at least every six months.
01 02 03 04 05 06

Module 03 of 06

Build your first assistant with Projects

Demo 1 · Build a specialist

Build a specialist, then put it to work.

Now that Claude knows you, give it one job. The same idea, getting more specific each step.

Knows you

Master Prompt

General context about who you are and how you work.

general
One job

System Prompt

A specialist built for one task, dropped into a Project.

specific
Delivers

Real task

Finished deliverables, ready to use.

result

Demo 1 · The System Prompt

Give it real expertise.

You don't have to be the expert. Claude researches the field first, then writes the brief that makes it one.

A System Prompt is the brief for one job: the frameworks it follows, the method it uses, and what a good result looks like.

Frameworks Method Guardrails Output format Voice

Built with Deep Research

01Define the one job
02Deep Research the ideal expert
03Claude writes the System Prompt
04Drop it in a Project, then run the task
Claude does the research, so you don't have to be the domain expert.

Demo 1 · Two kinds of instruction

Master Prompt vs System Prompt.

Everywhere

Master Prompt

About you.

Lives in your profile. Applies to every chat you start.

One project

System Prompt

About one job.

It is the project's instructions. Applies inside that Project only.

Your turn

Okay, now you build it.

Build your own assistant from scratch. You have about 20 minutes, and you can ask questions any time as you go.

01Define the job that needs to be done
02If it needs it, do the deep research first
03Ask Claude to write the system prompt
04Drop it into a Project
05Add any resources it needs
06Test your assistant and iterate
About 20 minutes · ask questions any time
01 02 03 04 05 06

Module 04 of 06

Turn your assistant into a Skill

Module 04

What is a skill?

A skill is just a folder you can zip up and share (its packaged file ends in .skill). Plain files tell Claude how to do one job well, and it only opens them when it needs them.

The shape

your-skill/
SKILL.mdthe instructions (its system prompt)
scripts/optional code it can run
references/material to consult
templates/reusable templates and assets

An example

dashboard-builder.skill/
SKILL.mdThe recipe Claude follows: ask which numbers matter, read the data file, apply our colors, then build and lay out the charts. Its description, "build me a dashboard," is what makes Claude reach for it.
scripts/build_dashboard.py turns a spreadsheet into finished charts, so Claude runs the code instead of rebuilding the logic each time.
references/brand-guidelines.md holds our palette, fonts and chart styles, so every dashboard comes out on-brand.
templates/past-dashboards/ are two or three we shipped before, for Claude to match the layout.

Same four parts as any skill, filled with real contents. Then you invoke it anywhere: any chat, Cowork, or Code, by slash command or just by asking.

Project vs Skill

Same ingredients. One stays put, one travels.

In a Project
As a Skill
Instructions
Your system prompt
=
SKILL.md
Reference material
Files you add
=
references/
Templates & examples
Files you add
=
templates/
Code it can run
not its job
+
scripts/

Static. A project lives in one place. You bring your work to it.

Portable. A skill goes anywhere. Invoke it by slash command or just by asking.

Built an assistant you rely on? Because the anatomy is almost identical, you can lift it into a skill, use it everywhere, and share it with your team.

Module 04

When and why to make a skill.

Make a skill when a task is repeatable. Build it once and it works for you every time, and you can share it with the whole team so everyone runs your best process.

Two ways to build one

1  Plan it first

Scope what it should do, then write the SKILL.md and add the parts it needs. Good when you already know the process cold.

2  Do it, then compound it

Just do the task with Claude, refine until the output is exactly right, then say "turn everything we just did into a reusable skill." Often the easier path.

Your turn

Now make a skill.

Turn the assistant you just built into a skill, or create a brand-new one. About 10 minutes, and ask questions any time.

Easiest path: do the task with Claude, get the output right, then say "turn this into a skill."

About 10 minutes · ask questions any time
01 02 03 04 05 06

Module 05 of 06

A look at what's next

Demo 2 · Claude Cowork

Now bring Claude to your files.

The wall in Chat

Claude.ai is capped on file size and cannot reach your computer. You upload one file at a time.

What Cowork changes

Claude works directly inside a folder on your machine: every file at once, no uploads, no size limits.

Claude working inside a folder on your computer

Demo 2 · What we start with

A folder full of spreadsheets.

Folder of files
Customer returns log
Keyword rank tracker
Commercial invoice
Inventory snapshot
Amazon PPC report

Demo 2 · The prompt

Anatomy of a good prompt.

The dashboard is only as good as the brief. Here is what we gave Claude, and why each piece matters.

"Hey Claude, I need your help with something."

Context · who it is for

A folder, Harbor & Pine, for a client: an Amazon FBA seller buried in disconnected Excel files.

The problem · why it matters

He has no single place to see the state of his business or make decisions. It is cumbersome and easy to lose track.

The ask · what to do

Research the best dashboard designs and practices, read all of the data, and propose a highly visual, interactive dashboard he can act on.

The bar · what good looks like

A self-contained HTML file, easy to share and understand, following strong UX/UI and front-end practices.

Room to think · trust its judgment

Use HTML, CSS, React, or whatever fits. Think about what would truly be impactful for his business, and why.

Five parts, one complete brief: context, problem, ask, standard, and room to think.

Demo 2 · The result

Messy files become a live dashboard.

Harbor and Pine operations dashboard
Click to open the live dashboard

Demo 3 · Claude Code

Now hand Claude the keys.

The edge of Cowork

Cowork builds inside your folder. Putting things live, wiring up tools, and running big multi-part jobs is still on you.

What Code unlocks

Claude gets your whole machine and the internet: sub-agents for bigger builds, and live websites it can deploy and maintain, once set up, with very few manual steps.

You still don't write code. You describe the outcome.

Claude Code with keys to the whole machine

Demo 3 · The prompt

Anatomy of a bigger prompt.

Same idea, bigger reach. This brief asks Claude to research, design, check its own work, and put the site live.

"Okay Claude, on my desktop there's a folder called Harbor & Pine Goods."

Context · where the data lives

A folder on my desktop with the data from an Amazon FBA seller.

The goal · what to build and why

A commercial website that gives them a web presence: tell their story, show their products beautifully, and link back to their Amazon listings.

The ask · the build steps

Find which products they sell, spin up a sub-agent to search the web for high-quality photos, then use your front-end design skill and study great comparable sites.

Quality loops · make it check itself

Before showing me: critique your own design, spin up a graphic-design agent to review it against the design skill and propose fixes, then navigate the site with Playwright to confirm it holds up.

Ship it · deploy live

Host it live on Vercel using my Cloudflare account. You build and deploy; guide me through the one manual step, pointing the CNAME.

Same anatomy, bigger reach: now it researches, self-critiques, and ships.

Demo 3 · The result

The same data becomes a live website.

Harbor and Pine Goods website
Click to open the live website

Going further

A plugin is a box of skills.

A Skill

One capability: a SKILL.md folder of instructions. Claude reaches for it on its own when the task matches its description.

A Plugin

A bundle: many skills plus hooks, connectors and commands. Why bundle? Install it once, version it, and roll it out to your whole team, so everyone gets the same vetted toolkit, not loose files to copy.

Can a skill use another skill? Yes. A skill's instructions can tell Claude to run another skill, so they chain: a report skill that calls a graphic-design skill, or our call-triage skill that hands off to the summary skill. Claude does the calling.

Going further

Routines: set it, walk away.

A routine is a saved Claude session (a prompt, your files, your connectors) that runs on Anthropic's cloud, automatically, so it works even with your laptop closed.

On a schedule On an API call On a GitHub event
01Build the session you want
02Save it as a routine
03Attach a trigger: schedule, API, or GitHub
04It runs in the cloud; watch the run history

Going further

What to automate, and what not to.

Good fits
Morning briefing: overnight changes in a 5-minute digest
Inbox or backlog triage: label, assign, summarize to Slack
A weekly report or a docs-drift scan
A PR review digest on every new pull request

Recurring, self-contained, and reviewed later.

Not for
One-off, ad-hoc tasks
High-stakes calls needing real-time judgment
Anything needing your approval each step (routines run on their own)
Sub-hourly cadence (one-hour minimum)

A "green" run means it ran, not that it nailed it: spot-check.

01 02 03 04 05 06

Module 06 of 06

Other models worth knowing

Other models worth knowing

Claude's home base. Call a specialist when it helps.

Gemini
Google
For what's current

Deep Research grounded in Google Search and YouTube. Best when the answer changes week to week.

Grok
xAI
For the social pulse

Reads live posts on X. Best for breaking news, sentiment, and the talk of the town.

ChatGPT
OpenAI
For images

gpt-image-2 renders legible text and keeps characters consistent. Google's Nano Banana Pro is neck-and-neck.

Claude stays home base for analysis, synthesis, and building. These are specialists you reach for on purpose.

A research move

Two researchers beat one.

Run the same deep research on Claude and Gemini, then let Claude fuse the two.

The move

01Tell Claude the objective
02Claude writes the deep-research prompt
03Run it on Claude and Gemini
04Fuse both back in Claude
05A triangulated answer

What the fuse reveals

Shared truth
what both agree on, trust it most
Complementary
what only one of them found
Contradictory
where they disagree, dig in there

Another tool worth knowing

NotebookLM: it only knows what you give it.

Upload your own sources and it answers only from them, citing the exact passage, so it won't invent anything beyond your material. Great for studying a dense pile of documents or pressure-testing a dataset. Free with a Google account.

Careful with proprietary data. On the free tier your feedback can be human-reviewed. Use a Workspace or enterprise account for anything confidential.

Working with data

Trust, then verify.

Models sound most confident exactly when they're wrong. A few habits keep you safe:

Cross-reference — run the same task on another model and compare.
Spot-check — hand-verify a random sample; for a generated spreadsheet, check 5–10 cells against the source.
Ask for sources — "where did you get this?" and make it cite.
Run it a few times — keep the answer that recurs.
Human in the loop on high-stakes numbers.

A tip to take with you

Talk to it. Don't type.

Why dictate

About 3× faster than typing, and you give a fuller, richer prompt instead of rationing keystrokes. Wispr Flow: press a key, speak naturally, clean text lands in any app.

Why it works

These models are built from language: they learned by predicting the next word across enormous text. Natural, complete, spoken language looks like what they trained on, so it lands better than clipped typing.

A tip to take with you

Treat it like a smart colleague.

The newer models reward collaboration over commands. Don't bark a task, work with it:

Co-define the task first — share your standards and a few files, ask it to help shape the work, then say "go."
Think out loud together — it stays with you as you move from exploring to executing.
Brief it in plain language, the way you'd brief a coworker.
Then delegate real, meaty work.

Recap

One Claude. Three ways to work.

1Chat

Rent a cubicle

Claude.ai in your browser or on your phone. Best for everyday thinking: writing, analysis, research, and building a specialist inside a Project.

You give it a Master Prompt

2Cowork

Bring it to your folder

Works directly inside a folder on your computer. Best for turning a pile of files into something useful, like a live dashboard.

You give it Folder instructions

3Code

Keys to the machine

Your whole machine and the internet. Best for big builds Claude can research, refine through its own loops, and ship live.

You give it a CLAUDE.md

More reach at every step. The unlock never changes: give Claude context, and tell it what you want.

Thank you.

Now let's put it to work.

AI Sherpa

aisherpa.me  ·  cal.com/santiago-restrepo