We've all been there. You build the perfect browser automation script. It clicks all the right buttons, fills all the right fields, handles edge cases beautifully. Then the website updates—and everything breaks.
This is the fundamental flaw in how we've been teaching AI agents to interact with the web. We've been describing how to do things, not what needs to be done.
The Selector Trap
Traditional automation relies on brittle selectors:
await page.click('#submit-btn');
await page.fill('#username', 'john');
When #submit-btn becomes #submit-button or the class names change, your automation fails. You're not describing the goal ("submit the form"), you're prescribing the implementation ("click element with ID submit-btn").
This creates an endless cycle of:
- Write automation script
- Website updates
- Script breaks
- Debug and fix
- Repeat
A Better Way: Intent-Based Skills
SkillForge takes a different approach. Instead of capturing selectors, it captures intent.
Here's how it works:
- Record yourself performing any web task
- Extract the workflow using AI analysis
- Generate a SKILL.md file describing goals and context
- Execute with any compatible agent
The generated skill looks like this:
# Book a Meeting
## Goal
Schedule a meeting on Google Calendar
## Workflow
1. Navigate to calendar.google.com
2. Click "Create" button
3. Fill in meeting details
4. Add attendees
5. Send invitation
## Context
- "Create" button: Primary CTA, top-left
- Title field: Text input labeled "Add title"
Notice: no CSS selectors, no XPath, no brittle identifiers. Just semantic descriptions that an AI agent can interpret and adapt.
Why This Changes Everything
Resilience: When the UI changes, the agent adapts because it understands the goal, not just the mechanics.
Accessibility: Domain experts can create automation without learning to code. Just perform the task once.
Portability: Skills work across frameworks—AutoGen, LangChain, CrewAI, or custom implementations.
Auditability: Human-readable format means anyone can review what an agent will do.
Real-World Applications
- Sales teams: Record CRM workflows once, automate forever
- Operations: Document processes through demonstration
- Customer support: Create troubleshooting flows
- QA testing: Generate tests from user sessions
The Future of Agent Capabilities
We're moving toward a world where AI agents learn by observation, just like humans. The ability to teach through demonstration democratizes automation and makes it accessible to everyone.
No more brittle selectors. No more broken scripts. Just show the AI what you want, and it learns.
Check out SkillForge:
🔗 https://www.producthunt.com/products/skillforge-2
What will you automate when brittle selectors aren't a concern?
Top comments (0)