agentleFS
d4vidc4rsond4vidc4rson/paralogy-divergent-thinking-tools/skills/anti-homogeneity-check/SKILL.md

anti-homogeneity-check

Skill5 starsChanged 8 months agoMITView on GitHub

Post-generation diversity audit that scores whether a set of ideas are actually different from each other or just surface variations of the same thought. Use AFTER any brainstorming, ideation, or creative generation — from any source, any skill, any person — when you need to know if your "ten different ideas" are actually ten ideas or three ideas wearing costumes. Triggers on "are these actually different," "these all feel the same," "did I just write the same idea five times," "check these for diversity," "which of these overlap," or any moment after generating a list where the user wants to know what they actually have before moving forward. Also use automatically as a quality gate after any other creative skill produces output, or when a user pastes in a list of ideas from anywhere and wants them audited. Works on any problem in any domain — strategy, engineering, science, policy, design, education, medicine, law, research. This is the bullshit detector for brainstorms.

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 22 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈22

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Not what the idea says on the surface.

    line 75
  2. Distinctive:

    Only the vocabulary changes.

    line 79
  3. Distinctive:

    No cluster contains more than 30% of the ideas.

    line 216
  4. Distinctive:

    Not by generating new ideas — by identifying the empty spaces.

    line 287
  5. Distinctive:

    Look at the clusters you found and ask: what kind of move is completely absent from this set?

    line 292
  6. Distinctive:

    All digital, nothing physical?

    line 305
  7. Distinctive:

    All intervention, nothing structural?

    line 306
  8. Distinctive:

    All about acquiring new people/resources, nothing about deepening what already exists?

    line 307
  9. Distinctive:

    All communication, nothing about systems?

    line 309
  10. Distinctive:

    Do all the ideas target the same group?

    line 312
  11. Distinctive:

    Don't generate ideas to fill them.

    line 329
  12. Distinctive:

    Every idea here is about reaching new people.

    line 335
  13. Distinctive:

    Every idea here operates through the same channel.

    line 338
  14. Distinctive:

    Every approach here screens for the same thing using different methods.

    line 348
  15. Distinctive:

    When the user has already decided and is in execution mode

    line 383
  16. Distinctive:

    When the user explicitly wants deep variations on a single theme (deliberate clustering is fine — accidental clustering is the problem)

    line 385
  17. Distinctive:

    Use other skills for quality assessment.

    line 404
  18. Distinctive:

    Use this one to see clearly before you choose.

    line 405
  19. Distinctive:

    Run it after the Guilford Engine's final output to catch clustering the pair-check missed.

    line 416
  20. Distinctive:

    If the Anti-Homogeneity Check gives a Grade C to Persona Engine output, the personas weren't different enough.

    line 423
  21. Distinctive:

    If 30 ideas sort into 4 clusters, you have 4 ideas to evaluate, not 30.

    line 433
  22. Distinctive:

    If 6 of the 10 are in the same cluster, the user needs to know that before they accidentally pick three that are all the same move.

    line 440

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Anti-Homogeneity Check

You just generated ten ideas. How many ideas do you actually have?

The research says: probably three.

Doshi & Hauser (2024) proved that AI-assisted creative output is rated more creative individually but is significantly more similar collectively. Wenger & Kenett (2025) confirmed that LLMs show high creative homogeneity even across different model families. Ideas from a single AI perspective cluster at 0.92 cosine similarity.

That means most brainstorms — whether from AI, from a team meeting, or from a solo session at midnight — produce the illusion of variety while actually circling the same center.

This happens in every domain. Marketing brainstorms produce ten ideas that are all "get attention through someone else's audience." Engineering brainstorms produce ten ideas that are all "add another layer of abstraction." Research brainstorms produce ten ideas that are all "run the same experiment with a larger sample." Policy brainstorms produce ten ideas that are all "expand the existing program." Education brainstorms produce ten ideas that are all "make the same lesson more engaging."

The surface looks different. The skeleton is the same.

This skill is the x-ray. It shows you the skeleton underneath the ten ideas that look different on the surface, so you can see how many bones you're actually working with.


How It Works

Feed this skill any list of ideas. From any source. Any format. Any domain.

It does three things:

  1. Cluster — Groups the ideas by what they actually share underneath
  2. Grade — Scores the overall diversity of the set
  3. Diagnose — Tells you what's missing

It does not generate new ideas. It does not fix the list. It tells you what you have so you can decide what to do about it.


Step 1: Cluster

For every idea in the list, ask: "What underlying move is this making?"

Not what the idea says on the surface. What it's doing underneath.

The clustering logic is the same in every domain. Only the vocabulary changes.

In marketing, these look like three different ideas:

  • "Launch a TikTok series"
  • "Partner with influencers"
  • "Build a social media community"

Underneath, they're all making the same move: "Get attention through other people's audiences online." One idea. One cluster. Three costumes.

In engineering, these look like three different ideas:

  • "Migrate to microservices"
  • "Implement event-driven architecture"
  • "Decompose into bounded contexts"

Underneath, they're all: "Break the monolith into smaller pieces." One idea. Three architecture diagrams.

In science, these look like three different ideas:

  • "Screen a larger compound library"
  • "Apply machine learning to predict candidates"
  • "Use high-throughput assays"

Underneath, they're all: "Search faster through the same solution space." One idea. Three methods.

In policy, these look like three different ideas:

  • "Increase funding for the program"
  • "Expand eligibility requirements"
  • "Scale the pilot to more districts"

Underneath, they're all: "Do the current thing bigger." One idea. Three budget line items.

Conversely, these two might look similar:

  • "Host a dinner for 20 people"
  • "Send handwritten letters to 200 people"

Underneath, they're making different moves: The dinner creates intimacy through shared experience. The letters create intimacy through personal effort at scale. Different clusters. Similar surface, different bones.

How to name the cluster

The cluster name is the move, not the tactic. State it as a verb phrase. What is this idea doing?

Good cluster names:

  • "Get attention through borrowed audiences"
  • "Search faster through the existing solution space"
  • "Break complex systems into smaller independent parts"
  • "Replace the ask with an invitation"
  • "Use scarcity to manufacture desire"
  • "Change what we measure rather than what we do"
  • "Remove a step instead of improving it"

Bad cluster names (too vague):

  • "Marketing ideas"
  • "Technical improvements"
  • "Policy interventions"
  • "Engagement"

Bad cluster names (too specific):

  • "TikTok ideas"
  • "Microservices ideas"
  • "Grant-funded approaches"

The cluster name should be specific enough that you can immediately tell whether a new idea belongs in it or not, but abstract enough that it describes the move rather than the medium.

Present the clusters

Group every idea under its cluster name. Show the user what they actually have.

Example output (nonprofit context):

Cluster 1: "Borrow credibility from trusted voices" (3 ideas)

  • Partner with podcasters for sponsored segments
  • Get featured in local newspaper profiles
  • Have existing customers write testimonials

Cluster 2: "Create physical artifacts of belonging" (2 ideas)

  • Design membership tokens
  • Install trail markers with donor names

Cluster 3: "Embed in existing routines" (2 ideas)

  • Partner with local businesses for member perks
  • Include membership in real estate closing costs

Cluster 4: "Make knowledge the status signal" (1 idea)

  • Build interactive Water Census map

Cluster 5: "Replace charity framing with club framing" (1 idea)

  • $700 membership, word "donate" disappears

Example output (research context):

Cluster 1: "Search the existing space faster" (4 ideas)

  • High-throughput screening
  • ML-predicted compound prioritization
  • Larger sample sizes
  • Automated assay pipelines

Cluster 2: "Redefine what we're looking for" (1 idea)

  • Measure population-level genomic shifts instead of individual survival

Cluster 3: "Change who's looking" (1 idea)

  • Collaborate with ecologists who study non-resistance adaptations

Now the user can see the truth: they thought they had six ideas. They actually have three moves, and one of those moves has four ideas in it that are all doing the same thing.


Step 2: Grade

Score the overall diversity of the set using the LiveIdeaBench fluency system, adapted for a full list rather than pairs.

The grading scale

Grade A — Genuine portfolio. 5+ distinct clusters. No cluster contains more than 30% of the ideas. The clusters address different problems, use different mechanisms, and would appeal to different stakeholders. This is a real brainstorm with real variety.

Grade B — Decent spread with one gravity well. 4+ distinct clusters, but one cluster is overweight (40%+ of ideas). There's variety, but there's also a clear center of gravity pulling the set toward one kind of thinking. Usable but lopsided.

Grade C — Illusion of variety. 2-3 clusters. Most ideas are doing the same thing in different clothes. This is what most brainstorms produce. It feels like you have options but you're actually choosing between flavors of the same answer.

Grade D — Monoculture. 1-2 clusters. Everything is the same idea. The "brainstorm" was really one thought exploring its own surface area. Start over.

Present the grade

State the grade. State the cluster count. State which cluster is heaviest.

Keep it blunt:

"Grade B. You have 5 clusters but 40% of your ideas are variations on 'borrow someone else's audience.' You have real variety in the other 60% but your center of gravity is pulling everything toward distribution tactics."

Or:

"Grade C. You have 8 ideas but they sort into 2 clusters. Six of them are different ways to 'search the same space faster.' Two of them are different ways to 'redefine what counts as a result.' You have two ideas, not eight."

Or:

"Grade C. You have 12 policy proposals but they sort into 3 clusters. Seven of them are different ways to 'expand the existing program.' Three are different ways to 'add enforcement mechanisms.' Two actually change the underlying model. You have three ideas, not twelve."


Step 3: Diagnose

Tell the user what's missing. Not by generating new ideas — by identifying the empty spaces.

Find the gaps

Look at the clusters you found and ask: what kind of move is completely absent from this set?

Common missing dimensions:

Time horizon gaps. Are all the ideas short-term? Is anything here a long play? Is anything here about next week?

Mechanism gaps. Is everything the same type of action? All digital, nothing physical? All intervention, nothing structural? All about acquiring new people/resources, nothing about deepening what already exists? All communication, nothing about systems?

Audience gaps. Do all the ideas target the same group? Is anyone being ignored who shouldn't be?

Risk gaps. Is everything safe and implementable? Where are the stretch ideas? Where are the moonshots? Or conversely — is everything ambitious with nothing you could do Tuesday?

Problem gaps. Are all the ideas solving the same sub-problem? What sub-problems does this list completely ignore?

Present the gaps

State 2-4 missing dimensions. Don't generate ideas to fill them. Just point at the empty space and let the user decide whether to fill it or not.

"What's missing: Every idea here is about reaching new people. Nothing addresses deepening relationships with people you already have. Every idea here operates through the same channel. Nothing creates a different kind of experience. There's no moonshot — everything could ship this quarter, which means nothing here changes the trajectory long-term."

Or:

"What's missing: Every approach here screens for the same thing using different methods. Nothing questions whether we're screening for the right thing. Nothing addresses the institutional incentives that keep the field running the same experiments. There's no idea for what we publish if the primary approach produces null results."


When to Use This Skill

Use it after:

  • Any brainstorm (from AI, from a team, from yourself)
  • Guilford Engine output (as a quality check on the fluency pass)
  • Persona Divergence Engine output (to verify the personas actually diverged)
  • Short Think output (to sort the ore before refining)
  • Any list of ideas from any source that the user wants audited before committing

Use it automatically when:

  • A brainstorm produces more than 8 ideas (the bigger the list, the more likely there's hidden clustering)
  • The user seems uncertain about whether their ideas are "different enough"
  • You're about to help the user narrow down a list to their top picks (clustering first prevents them from picking three ideas that are secretly one idea)

Don't use it:

  • On a single idea (nothing to compare)
  • When the user has already decided and is in execution mode
  • When the user explicitly wants deep variations on a single theme (deliberate clustering is fine — accidental clustering is the problem)

What This Skill Is Not

This is not a ranking tool. It doesn't tell you which ideas are best. It tells you which ideas are different.

An idea can be in a cluster of one and still be terrible. An idea can be in a cluster of five and still be the right answer.

Clustering is about variety, not quality. Use other skills for quality assessment. Use this one to see clearly before you choose.


Relationship to Other Skills

After the Guilford Engine: The Guilford Engine has its own fluency check, but it grades adjacent pairs. This skill grades the full set. Run it after the Guilford Engine's final output to catch clustering the pair-check missed.

After the Persona Divergence Engine: The whole point of the Persona Engine is that different personas produce different ideas. This skill verifies that they actually did. If the Anti-Homogeneity Check gives a Grade C to Persona Engine output, the personas weren't different enough.

After Short Think: Short Think produces 30 raw ideas with no evaluation. This skill is the first evaluation step: before you score for originality or feasibility, find out how many ideas you actually have. If 30 ideas sort into 4 clusters, you have 4 ideas to evaluate, not 30.

Before any narrowing decision: Whenever a user is about to pick their top 3 from a list of 10, run this first. If 6 of the 10 are in the same cluster, the user needs to know that before they accidentally pick three that are all the same move.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.