Create a multi-step assignment with forms and a safe ending. Return ONLY JSON: {"goal":"final goal","steps":"steps in order, including filling out forms and filters","fill_rules":"what data to fill out form fields and what to substitute when choosing filters (for example: address - your city and street; filters - vegetarian, up to 30 minutes, rating 4.5)","stop_conditions":"when to stop and what to do when stuck (captcha, registration, dead end)","output":"final summary format"}
Multi-step tasks break down at the junctions: the agent reaches the form, gets stuck on a captcha (the “I’m not a robot” check), stumbles upon a registration requirement - and the chain breaks. Mechanics: each step depends on the previous one, so the assignment should provide not only a “path according to a happy scenario,” but also forks in case of obstacles. Engineers' trick: write the steps like a checklist with stopping points - “if the site requires a login, skip it and move on to the next one,” “if a captcha appears, stop and report.” So one gag does not ruin the whole task. The second trick: for forms, specify in advance what to fill in the fields (fill_rules) and what to substitute when choosing filters - otherwise the agent will start guessing your address or preferences. And keep the ending safe: collect and show, but do not order or pay - irreversible only with confirmation. Insider: ask the agent to keep a short “journal of steps” - what he did, where he got stuck. If the result is strange, the log will immediately show at what step everything went wrong. A typical mistake is to set only the ideal route without emergency exits; At the first obstacle, the agent either freezes or improvises at random.
Unlock access to submit solutions for instant AI review.