The generator on this site is one feature: a prompt, a model, a template back. It's not a facilitator, it's not an oracle, and it's not the whole site. The other half — the activity library — is curated by humans. The two halves do different jobs and the difference matters.
You pick a team type, a language, a topic count, and an optional theme. Our backend builds a structured prompt, sends it to a current-generation hosted LLM, and returns a template you can run in TeamRetro, in Miro, on a wall, anywhere. The shape is narrow on purpose: themes and topic counts, not freeform agentic generation. Narrow shapes hallucinate less.
We don't pin a specific model name in this copy because models rotate. The provider is a major hosted-LLM vendor and the system uses prompt-engineering rather than fine-tuning.
It doesn't read your team. It doesn't know your last sprint went badly. It doesn't decide whether a retrospective is the right meeting. It produces a template — a structured prompt for the people in the room — and that's the limit. The facilitation is still on you.
LLMs invent things. The generator is best at themed variations on formats that already exist — a Halloween-themed Start/Stop/Continue, a sports-themed Sailboat, a sci-fi Mad/Sad/Glad. Ask for a novel facilitation framework and you'll often get a confident-sounding remix of the canonical ones with the names shuffled.
Treat the output as a draft. Read it before you put it in front of the team. If a column doesn't make sense for your context, cut it.
Post-incident retros. Conflict resolution. Contentious decisions. Anything where the format itself is doing psychological work — anonymity, structure, a known sequence — pick a curated format instead. The activity library exists for that. Pre-mortem, Sailboat, Start/Stop/Continue and the rest have been stress-tested across thousands of teams; they survive the high-stakes moment in a way an AI-generated template doesn't need to be tested for.
The AI is a fast warm-up tool for low-stakes weekly cadence. That's the job it's good at.
Your prompts go to our backend, get sent to the LLM provider, and the response gets stored under a slug so you can come back to the template. We don't train models on your prompts. The provider may cache prompts transiently as part of operating their service; we don't add to it. See the privacy page for the data flow.
Anything served from /:templateId is AI-generated. Anything under /activities/:slug is curated by a human editor with field notes and pitfalls. The two surfaces are kept separate so you always know which one you're reading.