Teaching Agents by Example Not Code

The best way to teach a human is through demonstration and practice. So why do we teach AI agents through code and configuration files?

It's time to rethink how we train AI agents.

The Code-First Problem

Traditional agent development looks like this:

1. Write detailed instructions
2. Define API endpoints
3. Specify selectors and element IDs
4. Handle error cases in code
5. Debug when things break

This approach treats agents like traditional software—something you build once and deploy. But agents need to learn, adapt, and evolve. They need training, not just programming.

Learning by Example

What if we taught agents the same way we teach people?

• Show them what to do
• Let them observe the patterns
• Give them opportunities to practice
• Provide feedback on their performance

This is learning by example, and it's how SkillForge approaches agent training.

The SKILL.md Approach

Instead of writing code, you record yourself performing a task. The AI watches, learns, and extracts the essential patterns:

# Process Expense Report

## Goal
Submit a monthly expense report

## Workflow
1. Navigate to expense portal
2. Click "New Report"
3. Add expenses with receipts
4. Submit for approval

## Context
- Portal accessed via company SSO
- "New Report" is primary CTA
- Receipts uploaded via drag-and-drop

Notice what's missing: No code. No selectors. No brittle identifiers.

Why Examples Beat Code

Accessibility: Domain experts can create training examples without engineering support.

Maintainability: When a website changes, the agent adapts because it understands the task, not the implementation.

Clarity: Human-readable SKILL.md files make it easy to review and validate what agents will do.

Transferability: Examples can be shared, versioned, and reused across projects.

The New Training Loop

Step 1: Demonstrate
Record yourself performing any web-based task.

Step 2: Extract
AI analyzes the recording and generates a SKILL.md file describing the workflow.

Step 3: Validate
Review the SKILL.md, add edge cases, clarify ambiguous steps.

Step 4: Deploy
Use the skill with any compatible agent framework.

Step 5: Iterate
As the agent encounters new scenarios, refine the examples.

Real-World Impact

A sales operations team needs agents that can process leads from multiple sources. Instead of writing complex integrations:

• Record the lead processing workflow once
• AI extracts the SKILL.md
• Deploy across the team
• Refine based on real usage

Setup time: 30 minutes instead of 2 weeks.

Live on Product Hunt

SkillForge makes this possible today:

🔗 https://www.producthunt.com/products/skillforge-2

🌐 https://skillforge.expert

Stop writing code to teach agents. Start showing them what you want.

What will you teach your agents by example?