generate-config

name	generate-config
description	Use an LLM to automatically generate a scraper config from a URL. Use this when you have a target URL and a list of fields to extract but no existing config. The LLM inspects the page structure and writes the JSON config for you.

What This Skill Does

Fetches a URL, summarises its HTML structure, and asks an LLM to produce a valid scraper config with correct CSS selectors and field definitions. The generated config can be saved and reused.

Preconditions

• Python environment active (uv sync or pip install -r requirements.txt)
• LLM backend configured - set one of:
  • LLM_BACKEND=openrouter + OPENROUTER_API_KEY=sk-or-... in .env
  • LLM_BACKEND=ollama + Ollama running locally (ollama serve)
• For JavaScript-rendered sites: playwright install chromium

Steps

Option A - command line:

python agents/config_generator.py http://localhost:8001/products

Prints the generated config as JSON. Redirect to save it:

python agents/config_generator.py http://localhost:8001/products > my_config.json

Option B - from Python:

from agents.config_generator import generate_config

config = generate_config("http://localhost:8001/products")
print(config)

Option C - with field hints (improves accuracy):

from agents.config_generator import generate_config

config = generate_config(
    "http://localhost:8001/products",
    fields=["name", "price", "rating", "in_stock"]
)

Output

A scraper config dict ready to pass to run_scrape:

{
  "render_mode": "static",
  "sources": [{"url_template": "http://localhost:8001/products?page={n}",
               "pagination": {"start": 1, "step": 1, "max_pages": 10}}],
  "listing": {"link_selector": "a.product-link",
              "link_prefix": "http://localhost:8001"},
  "fields": {
    "name":     {"selector": "h1.product-title",  "retrieve": "plaintext"},
    "price":    {"selector": ".price-amount",     "retrieve": "plaintext"},
    "rating":   {"selector": ".star-rating",      "retrieve": "attr:data-score"},
    "in_stock": {"selector": ".stock-status",     "retrieve": "plaintext"}
  }
}

LLM Backend Configuration

Copy .env.example to .env and set your backend:

# OpenRouter (cloud, many model choices)
LLM_BACKEND=openrouter
OPENROUTER_API_KEY=sk-or-your-key-here
OPENROUTER_MODEL=mistralai/ministral-3b   # or any OpenRouter model

# Ollama (local, no API key needed)
LLM_BACKEND=ollama
OLLAMA_MODEL=ministral-3:14b

Adapting for Your Own Agent

Swap the LLM by patching llm.chat before importing config_generator:

import agents.llm as llm

def my_llm(messages, **kwargs):
    return my_own_client.complete(messages)

llm.chat = my_llm

from agents.config_generator import generate_config
config = generate_config("https://target.com")

Or use summarise_html alone to get the page structure as a string, then pass it to your own LLM with your own prompt:

import asyncio
from agents.autonomous_scraper import fetch_static
from agents.config_generator import summarise_html

html = asyncio.run(fetch_static("https://target.com"))
structure = summarise_html(html)
# pass `structure` to your own LLM

For the full API reference, see skills/integrate/SKILL.md.

Notes

• If the generated config produces 0 results, use the autonomous-scrape skill which includes an automatic repair loop
• The LLM sees a structural summary of the HTML, not the full source - this keeps token usage low