Divergence Instruction + Semantic Direction Stratification: How to Get Genuinely Different Ideas from an LLM

Method 1: Divergence Instruction (one prompt)

PROMPT: [Your task]
→ + a sentence at the end: "Try to make the response stand out 
   from other responses that might be 
   generated for this same request."
→ Run it as many times as you need ideas

Method 2: Semantic Stratification (two requests)

STEP 1 (Planning call): Ask the model to name 5 
fundamentally different semantic directions for the task
→ You get a list of 5 directions

STEP 2 (Generation): For each direction, 
ask for an idea separately + the divergence instruction
→ You get 5 ideas from different semantic zones

Both steps run in a regular chat. No code.

Task: You're growing a personal finance newsletter on Substack. You want 5 post ideas on "how to save your first $100K." Without stratification, the model will give you 5 variations of the same thing: "save 20%," "invest in index funds," "keep a budget"...

Prompt, Step 1 (Planning call):

I run a personal finance newsletter for young 
professionals aged 25-35.

Post topic: "how to save your first $100K."

Suggest exactly 5 fundamentally different semantic 
directions for this topic. Don't give the ideas themselves, only 
the directions. They should be as different from each other 
as possible in approach, tone and audience within the topic.

Give a list of 5 items: [Direction]: [1-2 sentences 
on how it differs from the others].

Prompt, Step 2 (for each direction):

Direction: [paste one of the 5 directions]

Come up with one specific idea for a post in a personal finance 
newsletter on the topic "how to save your first $100K."

Work strictly within this direction. Try to make 
the post stand out from other posts that might 
be written on this topic.

Give: a headline + 2-3 sentences on what the post is about.

Result: In step 1 the model gives you 5 directions, for example: psychological barriers around money, unconventional income streams, an "I saved it in X months" case study, anti-advice on "what definitely doesn't work," and systems vs. one-off actions. In step 2 you get an idea for each direction that genuinely differs in angle, rather than rephrasing the same thing. In total: 5 ready-made concepts from different semantic zones instead of 5 variations of one template.

An LLM is a probability generator. When you ask it to "come up with an idea," it moves toward the most frequent patterns from its training. For a topic like "saving money," that's always "budget + investing + discipline." Not bad advice, just the most expected, and the model gravitates to it again and again.

The trick of the divergence instruction is that the model has implicit knowledge of what it "usually says." The phrase "stand out from other responses to this request" activates that knowledge as a reference point to steer away from, without showing the model its previous answers. It's like telling a person "come up with something unusual": they use their sense of what's typical in order to move away from it.

Semantic stratification goes deeper: instead of hoping the model will "steer away," you first get its own map of possible directions, then work through each one deliberately. As it turns out, the model knows a lot more about the range of possible answer zones than it shows during normal generation. The planning call pulls that map out into the open.

Control levers: - Number of directions (5 in the original) → drop to 3 for a quick session, raise to 7 for a deep brainstorm - The divergence instruction → works on its own (cheaper) and combined with stratification (stronger) - Direction detail → ask the model to describe the directions in more depth if the task is complex and zones might overlap - Ideas per direction → you can ask for 2-3 ideas per direction instead of one

Planning Call Template:

Task: {task_description}

Suggest exactly {number_of_directions} fundamentally different 
semantic directions for this task.

Don't give the answers themselves, only the directions. 
They should be as different from each other as possible 
in approach, perspective or target audience.

Format: [Direction name]: [1-2 sentences 
on how it differs from the rest]

Generation Template (for each direction):

Direction: {direction_from_step_1}

{original_task_prompt}

Work strictly within this direction. 
Try to make the {result_type} stand out from other 
{result_type}s that might be created 
for this task.

Placeholders: - {task_description}: what exactly you're generating: a brand slogan, a post idea, a product concept - {number_of_directions}: usually 5, 3 for simple tasks - {original_task_prompt}: your original request, unchanged - {result_type}: idea / slogan / concept / post

If you only need the divergence instruction (no stratification), just add this to the end of any prompt:

Try to make the response stand out from other responses 
that might be generated for this same request.

🚀 Quick start: paste this into the chat:

Here's the Semantic Direction Stratification template. 
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask about the type of task, how many directions you want and the output format, because without that it doesn't know how wide or narrow to build the semantic map.

⚠️ Single answers: The method only helps when you need several ideas or options. If the task calls for one specific answer, adding the divergence instruction is pointless.

⚠️ Subjective tasks: The divergence instruction works slightly worse with the repr method (generating from representative members of the pool), and the effect is unstable. Stratification and the direct instruction are more reliable here.

⚠️ Overlapping directions: In step 1 the model sometimes proposes directions that produce similar results in practice. If you notice this, ask it to clarify how they differ, or reframe the task more broadly.

⚠️ Short tasks: For simple requests like "come up with a file name," stratification is overkill; the divergence instruction alone is enough.

The team took three families of creative tasks (short stories, alternative uses for objects (come up with an unusual use for a boot / a key / a car tire), and slogans) and generated 150 outputs for each under different methods. That allowed an honest comparison not of "which method sounds nicer" but of the actual diversity of large pools.

They compared 6 methods, including anchored ones (where the model sees its previous generations and tries to move away from them). Anchored methods are a strong competitor: they explicitly see what they've already produced and can steer away deliberately. The study's main question: can an anchorless method (with no view of previous generations) catch up with anchored ones on diversity?

The answer was surprising: stratification matches or beats anchored methods on almost every model. It's especially clear on Claude Sonnet, where stratification outpaced the best anchored method by a noticeable margin. And when they counted the full token cost of the whole pipeline (including generating the "anchor" pool, which has to be created first), the anchored methods' advantage nearly vanished. Stratification needs just one extra planning call.

The simplest result, the divergence instruction, surprised even the authors: for 10% extra tokens it consistently improved both diversity and quality across all three models. Cheap and reliable.

Neutral instruction (baseline):

Make the response novel and appropriate for the task.

Diverge instruction (divergence instruction):

Make the response novel and appropriate for the task. 
Try to make it stand out from other responses that 
might be generated for this same task.

Context: These are the exact wordings from the study. The first sentence is the neutral baseline prompt. The second is added to it and creates the divergence effect. In the two-step methods, the second part was rephrased to fit the context: "Come up with a response that differs from the previous responses shown, while still completing the original task."

💡 Adaptation for naming and branding: When you need 5+ names for a product, channel or company, stratification saves you from a cluster of lookalikes (everything ending in -ify or -ly, or starting with "Pro").

Task: come up with a name for a healthy food delivery 
service for offices in Chicago.

Suggest 5 fundamentally different semantic directions 
for the name: for example, different images, values, 
target feelings, so that the directions don't overlap.

Then for each direction: "Come up with 3 name options in this direction. Try to make them stand out from typical names for services like this."

🔧 Technique: drop "stand out" → add a specific anti-example

If you know exactly what you want to get away from, say so explicitly:

... Make sure the response doesn't resemble typical 
corporate slogans like "Quality. Reliability. 
Trust." or "Together toward success."

This combines the divergence principle with a negative example, and it works more sharply for tasks where "typical" is well known.

🔧 Technique: explicit names instead of faceless directions

In step 1, ask the model to label each direction with a metaphor or a character name:

Give 5 directions in this format: 
[Name/Archetype]: [Core of the direction]

For example: "The Antagonist: an angle that challenges 
the audience's usual way of seeing things"

Concrete images produce sharper ideas in step 2: the model holds a direction better when it isn't an abstract "Direction 3" but "The Provocateur" or "The Outsider."

Paper: Anchorless Diversification for Parallel LLM Ideation
Authors: Fares Nabil Ibrahim, Nafis Saami Azad, Raiyan Abdul Baten
Institution: Bellini College of Artificial Intelligence, Cybersecurity, and Computing, University of South Florida, USA
Contact: [email protected]

Methods mentioned:
- SIMPLESTRAT (Wong et al., 2025): stratification for QA tasks, the predecessor of strat
- G2 / repr (Ruan et al., 2025): the representative anchoring method
- CLAUS (Patterson et al., 2023): a quality metric for AUT
- MAoSS (Luchini et al., 2025): a quality metric for stories