You've learned the theory. Now let's watch someone actually do it. Meet Sarah, an instructional designer who's about to build something to make her job easier. We'll follow her through the entire Plan → Build → Assess → Compound loop.
👤 Meet Sarah
Background: Sarah is a senior instructional designer at a corporate training company. She's been vibe-coding with Claude.ai for about a year, but recently learned about Claude Code and the compound engineering approach.
The Problem: Every time Sarah starts a new training project, she spends 2-3 hours manually creating course outlines from stakeholder interview notes. She has a template, but tailoring it to each project is tedious. She types the same structural questions, reformats content, and organizes topics—work that feels automatable.
The Opportunity: "What if I could feed Claude my interview notes and have it generate a structured outline that I can then refine?"
💡 Brainstorming: What Could I Build?
Before jumping to solutions, Sarah brainstorms what kinds of tools would make her life easier. She comes up with 15 ideas:
Sarah's Product Ideas
🎯 Choosing What to Build
Sarah looks at her list and thinks through each option:
"The Interview-to-Outline Generator hits my biggest pain point. I do this every single project, it takes forever, and the output is predictable enough that AI should handle it well. Plus, if I build this right, I can expand it later to include the Learning Objective Writer and Assessment Question Bank. It's a good starting point."
Why this choice?
- High frequency (every project uses it)
- Clear time savings (2-3 hours → 30 minutes)
- Straightforward input/output
- Room to grow
- She can test it immediately on her current project
🔄 The 4-Step Loop in Action
Step 1: PLAN
Sarah opens Claude.ai (yes, for planning!) and starts:
"I'm an instructional designer who wants to build a tool that converts stakeholder interview notes into structured course outlines. Before we start building anything, help me think through:
• What features would this need?
• What's the simplest version that would be useful?
• What file structure makes sense?
• What should I build first?"
After a 10-minute planning conversation, Sarah has:
✨ Sarah's Plan
MVP Features:
• Single HTML page with textarea for notes input
• "Generate Outline" button
• Output display area with formatted outline
• Copy-to-clipboard functionality
Later Enhancements:
• Save/load functionality
• Multiple outline templates
• Export to docx
File Structure:
• index.html - Main interface
• app.js - Logic for calling Claude API
• styles.css - Clean, professional styling
• Claude.md - Project context and prompts
Build Order:
1. Basic HTML structure
2. Claude API integration
3. Styling
4. Copy-to-clipboard
5. Test with real notes
Sarah documents this plan in a Google Doc. Now she's ready to build.
Step 2: BUILD
Sarah creates a project folder and opens Claude Code:
First, she creates the Claude.md file:
"Create a Claude.md file that explains this project. Here's what we're building: [pastes plan from Google Doc]"
Then she builds the interface:
"Following the plan in Claude.md, let's build the basic HTML structure first. Create index.html with:
• A clean header
• Textarea for interview notes (500px tall)
• Generate button
• Output area
• Professional but minimal styling inline for now"
Claude creates the file. Sarah opens it in her browser—it works! But she realizes the API integration is more complex than she thought.
"I need an API key. And I'm not sure about the security of putting it in client-side JavaScript. Let me pause and research this."
She goes back to Claude.ai to ask about API security. After learning about environment variables and backend options, she decides:
"For MVP, I'll use a simple approach: hardcode my API key for local use only, add the file to .gitignore. Later, if I want to share this tool, I'll add a proper backend."
Back in Claude Code, she documents this decision:
"Add a 'Decisions' section to Claude.md documenting that we're using a hardcoded API key for local MVP, will move to backend if sharing. Also add a .gitignore file that excludes any files with API keys."
Then she continues building:
"Now create app.js that:
• Sends the interview notes to Claude's API
• Uses this system prompt: 'You are an instructional design expert. Convert the following stakeholder interview notes into a structured course outline with modules, topics, and learning objectives.'
• Displays the response in the output area"
After some back-and-forth to handle errors and add a loading indicator, Sarah has a working prototype. Time elapsed: 90 minutes.
🔒 Security Pause: Before Testing with Real Data
Sarah is excited to test with real interview notes, but she pauses. Her company has strict data security policies. She needs to think this through.
"Wait. I'm about to send stakeholder interview notes to Claude's API. These notes contain client names, business strategies, budget discussions—confidential information. What are my options?"
Sarah reviews Anthropic's security documentation and talks to her IT team. Here's what she learns:
🔒 What Claude Code's API Offers
Data handling guarantees:
• API inputs are NOT used to train Claude models (explicit policy)
• Data retained for 30 days for trust & safety monitoring, then deleted
• SOC 2 Type II certified
• GDPR and CCPA compliant
• Encrypted in transit (HTTPS) and at rest
What it doesn't offer:
• Data never leaves your local network (it does—goes to Anthropic's cloud)
• Zero data retention (minimum 30 days)
• Self-hosted/air-gapped option (requires internet, cloud API)
Resources: trust.anthropic.com | Privacy Policy
Sarah's company has a Claude for Work Enterprise agreement, which includes a Data Processing Agreement (DPA). After reviewing with IT/Security, she gets approval with these conditions:
✨ Sarah's Security Approach
Approved use: Anonymized training needs analysis
Her process:
1. Remove identifiable information from interview notes before processing
2. Replace specifics with generic terms:
• "John Smith, VP of Sales" → "Stakeholder A, Sales Leadership"
• "Salesforce implementation" → "CRM implementation"
• "Q4 budget of $500K" → "Upcoming budget allocation"
3. Keep instructional content intact (learning needs, skills gaps, audience characteristics)
4. Use only for outline generation, not storage of actual confidential notes
What this means: The AI-powered outline generation works with anonymized needs analysis. Sarah can still save hours, but confidential business details stay secure.
⚡ Pro Tip for Your Workplace
Before building tools that process company data with AI APIs:
• Review your company's data classification policy
• Talk to IT/Security about what's acceptable
• Share Anthropic's trust documentation: trust.anthropic.com
• Consider anonymization as a standard practice
• If your company doesn't have Claude Enterprise, that might be worth requesting
With her security approach defined, Sarah is ready to test.
Step 3: ASSESS
Sarah tests it with anonymized interview notes from her current project:
"The output is... not quite right. The learning objectives are too vague, and it's not following our company's format. I need to refine the prompt."
She iterates on the system prompt:
"Update the system prompt to include:
• Use action verbs for learning objectives (Identify, Explain, Demonstrate)
• Format: Module → Topics → Learning Objectives
• Keep objectives specific and measurable
• Limit to 3-5 objectives per topic"
She tests again. Much better! But she notices another issue:
"It's generating outlines, but I can't easily copy them to paste into my project docs. The formatting gets lost."
"Add a 'Copy to Clipboard' button that preserves the markdown formatting when pasted."
After testing with three different sets of anonymized interview notes, Sarah is satisfied. The tool works and saves her significant time.
Step 4: COMPOUND
Before closing Claude Code, Sarah documents everything she learned:
"Update Claude.md with a 'Session 1 Notes' section documenting:
• What we built
• The API key decision and why
• The final system prompt that works well
• What we learned about formatting output
• Next steps: add save/load, try multiple templates"
She also creates a separate file:
"Create a README.md that explains:
• What this tool does
• How to use it
• Current limitations
• Future enhancement ideas"
✨ What Sarah Documented
In Claude.md:
Project purpose, architecture decisions, current system prompt, API key approach, what works, what doesn't
In README.md:
User guide, setup instructions, limitations, roadmap
In her own notes:
Time saved per project (~2 hours), improvements needed, other ID tools to build next
🚀 What Happened Next
Two weeks later, Sarah has a new training project. She opens her outline-generator folder:
"Claude, read the Claude.md file. I want to add a feature: let users choose between two outline formats—our standard format and a client-specific format."
Because of the documentation from Session 1:
- Claude immediately understands the project structure
- Claude knows where the system prompt is and how it works
- Claude suggests adding a dropdown that switches between two prompt variations
- Implementation takes 20 minutes instead of starting from scratch
By Session 5, Sarah has added:
- Save/load functionality
- Three outline templates
- Export to markdown
- A "refine outline" feature
Each session is faster than the last. The tool compounds.
📊 The Results
| Metric | Before | After |
|---|---|---|
| Time to create outline | 2-3 hours | 30 minutes |
| Outline consistency | Variable (depends on her energy) | Consistent (follows template every time) |
| Time investment | Ongoing (every project) | Front-loaded (build once, use forever) |
| ROI after 10 projects | - | ~20 hours saved |
💡 Key Takeaways from Sarah's Journey
What Made This Work
✏️ Now It's Your Turn
Think about your own work:
🎯 Your Project Planning
List 3-5 things you do regularly that take time and follow a pattern
Which task would save you the most time or frustration if automated?
Describe the simplest possible version that would still be useful
Write one paragraph about your problem and why you're choosing this solution
⚡ Pro Tip
Don't try to build something complex right away. Sarah's first version had four files and took 90 minutes. That's the sweet spot: useful, but not overwhelming.
📚 Resources & Further Reading
- Compound Engineering (Every.to) The original methodology that inspired this workflow
- Building with the Claude API If you want to build tools like Sarah's that call Claude
- Compound Engineering Plugin Every's open-source tools for this workflow
💭 Pause & Reflect
Before moving on, take a moment to consider:
- Which part of Sarah's process resonates most with you?
- What would you have done differently?
- How would you handle data security for the projects you want to build?
- Can you see yourself following this same workflow?
- What's the first thing you want to build?
🎉 Module 3 Complete!
You've learned the complete workflow. Now it's time to build your own real project.
Module 3 Complete • Up Next: Module 4 – Your First Real Project