Skill Forge is an extension built on Browser-act CLI. It lets AI agents explore a website once, then generate a reusable skill for repeated extraction or automation.
The problem
Without a reusable skill, an agent has to rediscover the same website every time:
paths change between runs
failure modes differ
scaling is unreliable
token usage grows with every call
Skill Forge separates exploration from reuse.
Note: Explore once. Generate a deployable skill package. Reuse the same stable path for hundreds or thousands of inputs.
Installation
Skill Forge is separate from the Browser-act entry Skill. Install the entry Skill and CLI first.
Ask your agent:
Install browser-act-skill-forge. Skill URL: https://github.com/browser-act/skills/tree/main/browser-act-skill-forge After installation, verify that it is available.
Requirements:
Browser-act entry Skill installed
A task that can be described in natural language
Four-step pipeline
flowchart LR Describe["Describe"] --> Explore["Explore"] Explore --> Generate["Generate"] Generate --> Test["Self-test"]
Step | What happens |
01. Describe | Tell the agent what data or action you need |
02. Explore | The agent looks for an API-first path, then falls back to DOM automation if needed |
03. Generate | The agent creates a parameterized skill package |
04. Self-test | The generated skill is tested and repaired until it works |
01. Describe the task
Use natural language:
Extract job title, company, salary, and URL from LinkedIn jobs. I will later run this for 300 keywords.
02. Explore API-first, DOM fallback
Skill Forge prefers stable paths:
Priority | Strategy | Stability |
1 | API-first through network capture | More stable when backend contracts stay the same |
2 | DOM fallback through browser automation | Works when no useful API path is available |
03. Generate a parameterized skill
Business variables become CLI parameters instead of hard-coded values:
job-search --keyword "designer" --location "Berlin" --limit 100
Typical output:
SKILL.mdexecutable script
usage instructions
self-test evidence
04. Self-test and repair
The generated skill should be tested end to end. When the target site changes, the agent can re-explore and update the skill instead of rediscovering the workflow on every run.
Business value
For individual developers
Pay the exploration cost once, reuse the generated path for many inputs, and re-explore only when the site changes.
For teams
Move repeatable scraping work from one-off scripts toward generated, tested, reusable capabilities.
Traditional model:
request -> engineer writes scraper -> test -> deploy -> maintain
Skill Forge model:
request -> agent explores -> skill generated -> self-test -> deploy
The engineering work shifts from writing every scraper by hand to maintaining a system that lets agents generate reusable capabilities.
Use cases
Batch data collection
Run one pattern across many keywords, regions, or categories.
Cross-site price comparison
Generate skills for multiple sites and normalize output.
Periodic monitoring
Track price, inventory, or listing changes.
Internal processes
Wrap repeated browser operations into reusable skills.
Entry Skill vs. Skill Forge
Area | Browser-act entry Skill | Skill Forge |
Role | Gives the agent Browser-act CLI capability | Generates reusable task-specific skills |
Required | Yes | Optional |
Usage | Agent operates the browser in real time | Agent explores once and reuses generated skill |
Token cost | Paid during each live exploration | Concentrated during skill generation |
Browser-act gives agents the ability to act. Skill Forge helps agents turn a successful action path into a reusable skill.
Learn more
Installation
Install the base CLI and entry Skill before using Skill Forge.
Anti-detection & Blocking
Use blocking tools during exploration when a site challenges automation.
Designed for Agents
Use network capture and semantic descriptions effectively.