Reduck
Reduck docs
Back to Reduck

Using Reduck

Writing scripts

When no saved script fits, your agent writes one. You do not write selectors; you ask, and it works the site out in a live browser:

Use Reduck to create a script for trends.google.com keywords.

This page is about what it does with that, and how to judge the result. The full reference the agent reads is one command away — npx @reduck-ai/cli@latest skill authoring_scripts — and it loads the same text through read_docs.

How it goes

  1. Open a session. start_session gives the agent a browser it keeps across calls, unlike a run which opens and closes its own.
  2. Explore. With exec_code it navigates, looks at the page, and finds what it can rely on. Most of the work is here.
  3. Save. create_script stores the body together with the input and output JSON Schemas — the contract callers get.
  4. Test. It runs the saved script for real and compares the result against the trace.
  5. Iterate. Fixes go into a draft version, tested by id, promoted when they hold.

What makes a good script

One affordance of the site, not one of your projects. A script names something the site can do: send a message, list slots, search. find_pain_points is a research project, and Reddit has no button for it. Sort, filter, date range and pagination are arguments to a script, not scripts of their own — so your agent composes them, and the catalogue stays finite.

Arguments a caller can type from memory. A city, a query, a profile handle. Never an internal token you could only get by driving the UI by hand.

Output that mirrors the page, absence included. A book with no spec table, a listing with no price, a page a country blocks — those are answers, not failures, and the schema keeps them as null rather than dropping the record.

Anchors that outlast a redesign. A script reads what the site’s own code depends on — its hydration state, its Schema.org markup, its accessibility roles, its URLs — rather than class names and button text. Visible labels are the worst of the lot: text="Connexions" works in a French Chrome and breaks in an English one, without saying why.

Loud failure. A script that catches its own errors and returns nothing is worse than one that crashes: the crash names the line where the assumption was wrong. If a script has to recover from its own actions, the fix belongs where it started, not in a retry.

What you can do

  • Ask for the trace, not the result, when something looks off. The trace shows the steps and the screenshots behind the number that surprised you.
  • Write down what the site does to you — a rate limit, a login that expires, a page that lies to signed-out visitors — as a host README. Every script on that site gets it, and so does the next agent that touches one.
  • Run it twice, on different days. A script passes in the session it was built in and fails on the second run more often than any other way. A second run in a clean browser is the real test.
  • Keep it private or share it. A new script is yours; move it to a project for your team, or make it public for everyone. See Scripts.

Type to search the documentation.