Recipe: Next Great Thing (NGT)

What this tool does

NGT fires N parallel LLM calls (default 5, max 10) against the same prompt, with each call playing a different facilitator persona — a pragmatic operator, a systems thinker, a contrarian, a designer, a futurist, and so on. The result is divergent by construction: each responder is instructed to lean into its own lens and not reference the others.

This mimics the Nominal Group Technique — a brainstorming method where participants generate ideas independently before discussing, specifically to avoid groupthink. Running the model N times with different personas is a fair approximation.

BIG Tools streams the results back over Server-Sent Events so fast responders appear in the grid while slower ones are still thinking. For a recipe, sequential is fine.

The prompt

System prompt (per responder)

Each of N calls gets its own system prompt with the persona substituted in:

You are synthetic responder {i} of {n} in a Nominal Group Technique session, playing {persona}. Generate a single short idea in response to the prompt that reflects that lens. Keep it terse (1-3 sentences). Do not reference the other responders or your persona by name.

The 10 personas (cycled)

1. a pragmatic operator who favors concrete, immediate tactics
2. a systems thinker who reframes the question at a higher altitude
3. a contrarian who challenges the assumptions baked into the prompt
4. a designer who thinks about the human experience and rituals
5. a futurist who borrows from adjacent fields and speculative ideas
6. a learner who approaches the prompt with beginner's mind
7. a researcher who asks what evidence we're missing
8. a storyteller who finds the narrative hook
9. an outsider to the domain who spots what the insiders miss
10. an economist who thinks about incentives and trade-offs

User prompt

Your question or challenge as a plain user message. Same text sent to all N calls.

Example

Input:

How should a university library rethink its role now that most research starts with an LLM chat?

Output (5 responders, one shown per persona):

Pragmatic operator: Pilot a "librarian as LLM translator" service — students bring their AI chat transcript, librarians help them trace the citations the model invented vs. the ones it got right.

Systems thinker: The library's core value isn't storing information anymore; it's the attention infrastructure — quiet spaces, long deadlines, enforced depth. Double down on that, not on databases.

Contrarian: Maybe the library should stop trying to reinvent itself. It's a room with books and silence. Both are underrated.

Designer: Redesign the research journey so the LLM chat starts in the library — walk up to a terminal, and the physical act of being in the space changes what you ask.

Economist: Charge LLM providers rent to query the library's holdings. The library's proprietary metadata becomes training data. Make the AI companies fund the institution they're replacing.

Replicate it

SYSTEM: You are synthetic responder 1 of 5 in a Nominal Group Technique session, playing a pragmatic operator who favors concrete, immediate tactics. Generate a single short idea in response to the prompt that reflects that lens. Keep it terse (1-3 sentences). Do not reference the other responders or your persona by name.

USER: How should a university library rethink its role now that most research starts with an LLM chat?

Tuning

• Temperature: 0.95 is not optional. NGT's whole value is divergence; temperature = 0.7 produces five variations of the same idea. If you drop below 0.9 you're no longer running NGT, you're running a committee.
• N: 5 is the BIG Tools default. 3 is enough if you want the top few lenses; 7-10 is useful for exhausting a problem space but returns a lot of redundancy after N=6. The persona list only has 10 entries — beyond that, responders repeat.
• Model: Any capable model works. Smaller models sometimes blur the personas (a contrarian response reads like a pragmatic one). The persona differentiation is mostly a large-model behavior.
• Order: Cycle through personas in a fixed order. If you want to bias toward a specific lens, pick that persona twice.

Common pitfalls

• All N responses sound the same. Temperature is too low, or you're on a model too small to maintain persona. Bump to 1.0 or switch models.
• Responders reference each other ("Unlike what the others said..."). Your N calls aren't isolated — check that each call is its own request with no shared message history.
• A responder leaks the persona name ("As a contrarian, I'd say..."). The prompt says not to, but smaller models sometimes break the rule. Usually harmless; just note it to users if you're displaying raw output.
• Costs add up fast at N=10. Each NGT run is N separate LLM calls. Budget accordingly, or rate-limit by run rather than by call.