Practical Exercises

"Don't read prompts, write them. Don't look at examples, repeat them."

All exercises should be done in your terminal. Save the results to the my-experiments/ folder.

📝 Exercise 1: Rewrite a command into a delegation

Goal: Learn to go from commands to describing the result.

Task:

You have this "prompt" (a command):

Install npm
Create package.json
Install React
Install Tailwind
Create a src folder
Create App.tsx

What to do:

  1. Rewrite this as a delegation (describing the result)
  2. Save to my-experiments/exercise-1-delegation.md

Example result:

# Exercise 1: Delegation

## Input (WHAT WAS):
Step-by-step commands...

## Output (WHAT BECAME):
I want to build a React app for [PURPOSE].

Requirements:
- TypeScript support
- Tailwind CSS
- Structure: components in src/

Result: an app that runs on localhost:3000

Check:

🔄 Exercise 2: Experimenting with prompts

Goal: Learn to iterate and compare results.

Task:

Create a button component. But do it in 3 iterations:

Attempt 1: Minimal prompt

Create a React button component.

Attempt 2: With details

Create a React button component.
- Supports primary/secondary variants
- Color changes on hover
- Disabled state

Attempt 3: Full prompt

Create a React Button component for a design system.

Requirements:
- Variants: primary (blue), secondary (gray), danger (red)
- Sizes: sm, md, lg
- States: normal, hover, active, disabled
- On hover: scale 1.05
- Disabled: opacity 0.5
- TypeScript, with prop typing

Result: a Button.tsx file with usage examples

What to do:

  1. Launch Claude Code in the folder my-experiments/button-exercise
  2. Try all 3 prompt variants
  3. Compare results
  4. Save to exercise-2-iterations.md

Save:

# Exercise 2: Iterations

## Attempt 1: Minimal prompt
[Result from Claude Code]

## Attempt 2: With details
[Result from Claude Code]

## Attempt 3: Full prompt
[Result from Claude Code]

## Analysis:
- Which variant was the most effective?
- How many follow-up iterations were needed?
- Which details were the most important?

Check:

📋 Exercise 3: Structuring context

Goal: Learn to create good CLAUDE.md and TODO.md files.

Task:

Take an existing project (anything on your disk). For it, create:

  1. CLAUDE.md — full project context
  2. TODO.md — current tasks
  3. AGENTS.md — instructions for Claude Code

CLAUDE.md structure:

# Project Context

## My role
[Who I am for this project]

## Project
[What it is, why it exists]

## Stack
[Tech and versions]

## Structure
[Folders and their purpose]

## How to run
[Commands to start]

## Current state
[What works, what doesn't]

## When to call Claude Code
[For what tasks]

What to do:

cd [your-project]
touch CLAUDE.md TODO.md AGENTS.md

# Fill in each file
# Save results to git
git add CLAUDE.md TODO.md AGENTS.md
git commit -m "Add context files for Claude Code"

Check:

🎯 Exercise 4: Using Explore

Goal: Learn to explore a codebase with Claude Code.

Task:

Run /explore in an existing project with different parameters:

cd [your-project]
claude-code .

Inside the session:

/explore --thoroughness quick
# What did you learn in 30 seconds?

/explore --thoroughness medium
# What's the architecture?

/explore --thoroughness very_thorough
# How is everything connected?

Questions for Explore:

Try asking:

What to do:

Save the results to exercise-4-explore.md:

# Exercise 4: Explore

## Quick exploration
[Results]

## Medium exploration
[Results]

## Very thorough exploration
[Results]

## Questions and answers
- "Which files handle auth?" → [Answer]
- "How does..."

Check:

🧪 Exercise 5: Meta-prompting

Goal: Learn to create prompts that create prompts.

Task:

Write a meta-prompt for your project. This is a prompt that tells Claude how to create prompts for YOUR project.

Example:

Create a prompt for analyzing code in my [PROJECT] project.

The prompt should:
1. Ask the agent to understand the architecture first
2. Then analyze a specific file
3. Suggest improvements with project context in mind
4. Provide code examples

Make sure the prompt is clear and uses info from CLAUDE.md

What to do:

  1. Write the meta-prompt
  2. Give it to Claude Code
  3. Get the resulting prompt
  4. Use that prompt on real work
  5. Save to exercise-5-meta-prompt.md

Result:

# Exercise 5: Meta-prompting

## My meta-prompt:
[What I asked to create]

## Resulting prompt from Claude:
[The prompt Claude generated]

## How I used it:
[What task I applied it to]

## Result:
[Did it work?]

🚀 Exercise 6: Parallel agents

Goal: Learn to run several tasks at once.

Task:

Take a file with lots of info (for example, a list of user insights or a features backlog).

Launch 3 Claude Code sessions in parallel (in different terminals or folders):

Session 1: Analyzes the top 3 problems

From insights.md find and analyze the top 3 main user problems.
For each:
- Problem description
- How many users are affected
- A possible solution

Session 2: Finds the top 3 opportunities

From insights.md find the top 3 opportunities for new functionality.
For each:
- Opportunity description
- Which problem it solves
- Rough implementation difficulty

Session 3: Extracts the best quotes

From interview.md find the 5–7 most interesting user quotes.
The ones that reflect real pain points or important insights.

What to do:

# Terminal 1
cd ~/my-Efforts/On/analysis-1
touch insights.md
claude-code .
# [Give session 1 task]

# Terminal 2
cd ~/my-Efforts/On/analysis-2
touch insights.md
claude-code .
# [Give session 2 task]

# Terminal 3
cd ~/my-Efforts/On/analysis-3
touch interview.md
claude-code .
# [Give session 3 task]

# Combine the results into one file

Result:

Create presentation.md — a Markdown presentation for a stakeholder meeting:

# Product Analysis Presentation

## 🔴 Top 3 User Problems

### Problem 1
[From session 1]

### Problem 2
[...]

### Problem 3
[...]

## 🟢 Top 3 Product Opportunities

### Opportunity 1
[...]

## 💬 Best quotes

### Quote 1
> "[Quote]"
> — [Who said it]

[...]

## Recommendations
[Your analysis, what to do with this information]

Check:

📝 Exercise 7: Your own prompt template

Goal: Create a template you'll use for all your tasks.

Task:

Run through exercises 1–6, then create your ideal prompt template for your type of work.

It'll look like:

# My Perfect Prompt Template

## For backend development:
[Your template]

## For frontend development:
[Your template]

## For analysis and research:
[Your template]

## For documentation:
[Your template]

What to do:

  1. Recall which prompts worked best
  2. Which details were important?
  3. In what order do you give information?
  4. Which instructions do you always include?

Result:

Save to exercise-7-my-template.md

# My Prompt Templates

## Backend Development

## Role
You are a seasoned backend developer...

## Task
[Template for backend tasks]

## Input
[What info I provide]

## Output
[What I expect to receive]

## Constraints
[My constraints]

Check:

📚 Exercise 8: Documenting your style

Goal: Create a personal guide — exactly how you work with agents.

Task:

After completing all the exercises, write:

# My Working Style Guide

## When I delegate
[My criteria]

## When I decide myself
[My criteria]

## My ideal prompt contains:
[List of required elements]

## Mistakes I made:
[What didn't work and why]

## What worked best:
[Successful patterns]

## My work process:
[How I usually work with Claude Code]

## Tools I use:
[CLAUDE.md, TODO.md, AGENTS.md and others]

## How I track progress:
[How I know it's done right]

🎁 Bonus exercise: Build one automation

Goal: Go from a one-off prompt to a repeating pipeline.

Task

Open my-templates/automation-recipes.md and pick one recipe:

  1. Content Generation + Auto-Post to social media
  2. Auto-Reply Email (drafts in Drafts)
  3. Personalized Cold Email Outreach

Build it in an evening. Doesn't have to be perfect — the point is it works end-to-end.

Check

If >2 h/week — keep it and iterate. If not — delete without regret and try another.

🎨 Exercise 9: Vibe Coding — MVP in an evening

Goal: Run the full cycle from idea to working product using the AI stack.

What Vibe Coding is

An approach where you focus on product thinking, not technical implementation. AI tools cover the code, you set the direction.

Stack: Perplexity → Grok → Lovable → Supabase

ToolRoleWhat it does
PerplexityResearchMarket analysis, competitors, target audience
GrokPlanningStep-by-step instructions, prompt generation
Lovable / bolt.newFrontendFull UI from a prompt (React + Tailwind)
SupabaseBackendDatabase, auth, API without code

Task: Build a tool for yourself

Pick a task from your own life (not abstract!). Examples:

Steps:

Step 1: Research (Perplexity, 20 min)

Query: "What already exists in the category [YOUR IDEA]?
Top 5 competitors, their weaknesses, unmet audience needs."

Step 2: Plan (Grok, 15 min)

Query: "I want to build [IDEA]. Give me a step-by-step MVP plan.
Compose a prompt for Lovable/bolt.new — exactly what to write to it."

Step 3: Build Frontend (Lovable / bolt.new, 30 min)

Paste the prompt from step 2. Iterate — talk in design language:
"Make the button bigger", "Add a data table", "Add a sidebar menu"

Step 4: Connect Backend (Supabase, 20 min)

Result

Save to my-experiments/exercise-9-vibe-build.md:

# Exercise 9: Vibe Coding

## Idea: [What I was building]

## Research (Perplexity):
[What I learned about the market]

## Plan (Grok):
[Prompt I used to build]

## Frontend (Lovable / bolt.new):
[Screenshot or UI description]

## Backend (Supabase):
[Tables, what's saved]

## Outcome:
- How long it took: X hours
- Does it work end-to-end: yes / no
- Main insight:

Check:

✅ Checklist: All exercises

🎯 What's next?

After all the exercises:

  1. Practice daily — use these approaches in real work
  2. Record — save good prompts to my-templates/
  3. Share — show these exercises to coworkers
  4. Iterate — refine your style as you learn

📊 Final feedback: Deep JTBD Interview

After completing the exercises — fill in my-experiments/feedback-final.md (copy from my-templates/feedback-final-jtbd.md, 20–30 min).

This is deep reflection on the whole course through the 4 JTBD forces (Push / Pull / Anxieties / Habits) + before/after + ranking edits. Your signal (optional, anonymous) goes to course-feedback/ as a key input for the course's evolution.

Quick reference: CHEATSHEET.md

🎁 Track (optional): Package your expertise into a product

"Amplify your voice, don't replace it."

This track is for coaches, experts, practitioners, healers — people who already have a gift and want it to reach more people. It is not a "money machine" or a funnel with fake deadlines. It's about gathering and sharing what you already do well — in your own voice, with AI handling the packaging, not the substitution. Take one step at a time and save your work to my-experiments/.

Step 1. Name your gift

Goal: one paragraph, no sales copy.

Answer honestly: what do you give people, who exactly, what shift happens for them. Not a "sales pitch" — plain truth. Save to my-experiments/offer-1-gift.md.

Step 2. Find the recurring ask

Goal: find the core of your future product.

What question or pain do people bring to you again and again? That recurring ask is what's worth packaging first. Write down three such asks and pick one. Save to my-experiments/offer-2-core.md.

Step 3. Design it with your AI partner

Goal: structure where you lead.

Take the skill from Module 4 (prompt engineering) and ask the agent to help break your answer to that ask into a clear structure — a program, a guide, or a mini-course. You hold the meaning and the voice; the agent does the formatting. If it starts writing "how to sell" for you — stop it: you need structure, not manipulation. Save the draft to my-experiments/offer-3-design.md.

Step 4. Build a first version

Goal: a small, honest scope.

Vibe-code a first working version — a landing page, a PDF guide, or a single page with your offer. Not "perfect" and not "the whole program" — just enough for a real person to receive it and try it. Save the link/file to my-experiments/offer-4-build.md.

Step 5. Share, don't push

Goal: invite, don't pressure.

Show it to a few people it genuinely helps. No scarcity, no countdowns, no "2 spots left". Just: "I put this together, I think it'll be useful to you." Gather honest feedback and return to Step 3 — this is a loop, not a finish line. Note what you learned in my-experiments/offer-5-share.md.

Done. You didn't build a "funnel" — you let your gift reach people while staying yourself.

🔧 Track (optional): Document and automate your practice

"Take the routine off your hands — so there's time left for the human part."

This track is for anyone with a repeating practice: a coach, a maker, a freelancer, a small team. The goal is not to "replace yourself" or build a content mill — it's to lift the dull, repetitive part off you so you have time for what only you can do. AI takes the paperwork and the routine; the judgment and the human contact stay with you. Take one step at a time and save your work to my-experiments/.

Step 1. Pick one recurring task

Goal: one task, not the whole business.

Pick one thing from the menu that you do again and again:

Take exactly one — the one you're most tired of. Note your choice in my-experiments/sop-1-pick.md.

Step 2. Say it out loud — build the SOP with AI

Goal: turn "what's in your head" into clear steps.

Describe to the agent how you do this task — in your own words, the way you'd tell a colleague. Ask it to lay this out as a simple SOP: steps, inputs, what counts as "done". You edit and hold the meaning; the agent does the formatting. Save the SOP to my-experiments/sop-2-document.md.

Step 3. Find what you can hand to the machine

Goal: separate routine from judgment.

Walk through the SOP and mark it: where is mechanical routine (formatting, copy-paste, templates), and where does it need your judgment or a human touch. Automate the routine — leave the human part to the human. Write down exactly what you're handing to the machine in my-experiments/sop-3-automate.md.

Step 4. Build one small automation

Goal: one working piece, not the whole system.

Vibe-code or prompt one automation from Step 3 — an email template, a script stub, a checklist generator. Not "the whole process", just one honest piece that actually saves time. Test it on a real task. Save it to my-experiments/sop-4-build.md.

Step 5. Keep it alive

Goal: don't dump it in an archive.

Use the SOP and the automation on your next real task and fix what didn't fit. It's not a "finished product" but a living tool that grows with your practice. Note what you improved in my-experiments/sop-5-live.md.

Done. You didn't replace yourself — you cleared the routine out of the way so you can do what you're actually called for.

🔎 Track (optional): Research with AI

"Stop memorizing search operators — learn to research with AI."

"Knowing how to search" used to mean remembering Google operators. Now it means running a research loop with AI and verifying what it gives back. AI speeds up the search but can be confidently wrong and invent sources — so the rule of this track is simple: don’t trust the unsourced. Take one step at a time and save your work to my-experiments/.

Step 1. Pick a real question

Goal: a real question, not a "practice" one.

Take a question you'd google this week anyway — work, life, anything real. Phrase it in one sentence, the way you'd ask a live expert. Write the question down in my-experiments/research-1-question.md.

Step 2. Ask an AI search engine and run a follow-up loop

Goal: not the first answer, but a conversation.

Ask your question to an AI search engine — Perplexity, ChatGPT search, or Gemini. Don't stop at the first answer: refine, re-ask, ask it to narrow down or dig deeper. Good research is a loop, not a single query. Save your follow-ups and the key answers to my-experiments/research-2-loop.md.

Step 3. Use deep-research mode

Goal: depth where it matters.

For a question that genuinely matters, switch on a deep-research mode (deep research in ChatGPT, Gemini, or Perplexity). Give it time to go wide and assemble a report. Save what it found — together with its source list — to my-experiments/research-3-deep.md.

Step 4. Ground it and VERIFY the sources

Goal: separate truth from confident fiction.

Open the actual sources the AI cites and check the claims against them. If a link doesn't open, a source doesn't back the claim, or there's no source at all — strike that piece out. This is the heart of the track: don’t trust the unsourced, even when it sounds convincing. Note what held up and what didn't in my-experiments/research-4-verify.md.

Step 5. Synthesis: web search × AI

Goal: your own grounded answer, not someone's paraphrase.

Combine classic web search with AI synthesis: sometimes it's faster to open a couple of live pages yourself, sometimes to let the AI assemble the picture. Build your own short answer to the original question out of what you verified, and note where the AI got it wrong. Save the result to my-experiments/research-5-synthesis.md.

Done. You're not recalling search operators anymore — you run a research loop with AI and keep the sources in hand.