How to force AI to give you genuinely different options

Executive overview

When you ask AI for multiple options, it identifies its best answer first and then generates variations around that same core — different words, same idea. This is the gravity problem: all outputs orbit one answer.

Three prompt techniques break this pattern. A fourth method — subagents — eliminates shared context entirely.

The gravity problem

• AI picks its best answer first, then rotates variations around it
• Results look different on the surface but share the same structure and logic
• Affects emails, research, analysis, proposals, and strategic planning

Method 1: MECE constraint

• Add "mutually exclusive, collectively exhaustive" (MECE) to your prompt
• Forces outputs to cover the full problem space with no overlap between options
• Ask the AI to state in one sentence what makes each option different before writing it
• Works across any domain: writing, research, data analysis, planning

Method 2: Persona rotation

• Assign the AI distinct personas with fundamentally different worldviews
• Crucially: instruct the personas to disagree with each other — this is what guarantees divergence
• Ask the AI to name each persona and state its core belief upfront so you can validate the difference
• Example personas: minimalist (one core argument), analyst (data and facts), reframer (highlight the gap between current and future state)

Method 3: Dimension locking

• Keep most of the response fixed; change only one specific element per version
• Useful when you want to isolate the impact of a single variable (e.g., opening hook, argument structure, call to action)
• Tell the AI explicitly: differences must be meaningful, not just different words for the same idea

Bonus 1: Verification test

• After receiving options, paste this prompt: "For each version you just wrote, explain in one sentence what makes it fundamentally different from the others. If two versions share the same underlying idea, tell me."
• Forces the AI to grade its own output
• Gives you a one-sentence check per option to spot duplicates fast

Bonus 2: Subagents

• Available in Claude Code and Claude Cowork only
• A parent AI spawns separate child AIs (subagents), each with its own isolated context window
• No shared context means no gravitational pull between outputs — genuine independence
• Assign each subagent a distinct angle (e.g., growth operator, financial analyst, skeptical industry veteran)
• Provide rich context upfront: financials, client data, market research — or let subagents research online
• Ask the parent to summarise where subagents agreed and disagreed