Most people describing an automation don't name every tool involved. "Send a message when a new lead comes in" is a complete thought to the person typing it — but "send a message" could mean WhatsApp, SMS, Slack, or email, and guessing wrong doesn't just pick a slightly different template, it builds the wrong automation entirely.
Early testing surfaced a specific, repeatable failure: near-identical prompts that differed only in an unstated detail would resolve to completely different templates depending on which keyword happened to score highest in retrieval, even when neither prompt had actually specified a tool. The system wasn't wrong about matching words — it was answering a question the user never actually asked it to resolve on their behalf.
The fix isn't a smarter guess — it's recognizing when a guess is required at all. When a request leaves out a detail that materially changes which workflow gets built — which app, which channel, which notification method — Gydmation now asks a short clarifying question instead of silently picking one. A five-second follow-up question is cheaper than a wrong workflow that has to be found, understood, and rebuilt after the fact.
Not every gap deserves a question — asking about every minor ambiguity would make the product exhausting to use. We only interrupt for the specific class of missing detail that changes which nodes get used at all, not stylistic choices with a sensible default. Calibrating that line, so the product asks exactly the questions worth asking and no more, is ongoing work.
