method34

Your AI gives everyone the same idea.

EUREKA
BRIGHT SPARK
BRAINWAVE
BIG IDEA
AHA!
THE SPARK
GENIUS
BRILLIANT
NEW THINKING
BRAINSTORM
INSPIRATION
GOOD THINKING

Ask your AI for a metaphor about time. Odds are, it says river.

Write a metaphor about time.

Not just your AI: 25 models were asked this. Most said river or weaver.Artificial Hivemind, NeurIPS 2025

“Time” through method34:

Cream label, TIME IS A PEARL: an open hand drawn in fine engineering lines holds an open shell with a red pearl, a blue orbit ring around the pearl, red rays behind.

One pearl held 2,615 layers. It grew four or five a day for a year and a half. The oldest day is still buried in the deepest layers, smoothed over by every day since.

Method No. 3: Paired constraints
Mustard label, EVERY MOMENT IS MOLTEN GLASS: a flat geometric glassblower in a turquoise apron blows a large red ball of glass, a furnace with a red arch behind him.

The present is a blob of molten glass, soft enough to be shaped while it still glows. With every second it cools, a few more forms become impossible. And once cold, it is the past: its shape will never change again.

Method No. 33: Tree of ideas
Emerald label, TIME IS A SOURDOUGH STARTER: a flat geometric woman in a red headscarf kneads dough on a yellow board, a jar of bubbling starter beside her and a window behind.

Every day you let most of the sourdough go and give the last spoonful fresh flour. It rises again. Tonight’s bread comes from something alive since last spring, though barely a crumb of last spring’s flour is left.

Method No. 14: Diverge then converge
Blue and red retro poster label, TIME IS AN ELECTRIC CAR: an angular electric pickup on a tiled plain, mountains, an atom, a rocket and a red sun behind it.

When you lift your foot off the pedal and let a minute slow you down, the car quietly charges itself. Those slow minutes help you climb the next hill.

Method No. 27: Morphological box

This page came almost entirely out of method34. Except the method itself.

A match needs something rough to strike. AI needs this too.

Here are 34. Tap one.Hover over one.

Three people who never meet answer the same brief.

Method No. 01Persona setarXiv 2609.30492

Think it through in three languages; keep what only one found.

Method No. 06Language of thoughtarXiv 2601.11227

One property of an unrelated word shapes the answer, unnamed.

Method No. 15Random stimulusarXiv 2601.18053

Ban the default move and name what replaces it.

Method No. 03Paired constraintsarXiv 2607.27553

The model invents one odd person and answers only as them.

Method No. 02Meta-persona anchoringarXiv 2608.02618

Dice choose the cell; the model has to fill it.

Method No. 12Dice-conditioned generationarXiv 2610.01382

Force the answer into a rigid form and let the content bend.

Method No. 23Form constraintarXiv 2610.24223

A craftsperson and an expert from different worlds meet on one idea.

Method No. 07Maker + knower pairauthor's method

Ask the brief five ways; keep the answer that differs most.

Method No. 24Template rotationarXiv 2609.38829

Six thinking hats, one contribution each, one solution.

Method No. 20Parallel hatsarXiv 2503.12499

Ask what the task really is, name it in three words, solve there.

Method No. 04Reframe and name the frameWinston, MIT 6.034

Write the first three ideas, then forbid them.

Method No. 10Forbid the first NarXiv 2606.00875

Start from four mechanisms and deepen only the survivors.

Method No. 29Wide parallel initialisationarXiv 2610.00707

Map the field as a tree and join two leaves far apart.

Method No. 22Topic-tree distant pairingarXiv 2506.12317

A wild coach until the bell, then a strict one.

Method No. 21Diverge/converge coacharXiv 2510.26490

Design five worst ideas, then invert what makes them worst.

Method No. 26Worst possible ideaclassic

Ban every mechanism already used and look outside the list.

Method No. 09Seen-set exclusionWinston, non-redundant generator

List what professionals take for granted; break one by lot.

Method No. 25Assumption ledgerauthor's method

Generate ten, group by mechanism, keep one from the smallest group.

Method No. 13Scale then dedupearXiv 2607.27553

Fuse the core idea with a drawn domain into one that needs both.

Method No. 18Conceptual blendingarXiv 2106.14127

Five answers with their odds; keep the least likely one that works.

Method No. 11Verbalized samplingarXiv 2510.01171

Name the contradiction and solve it with inventive principles.

Method No. 19TRIZ contradictionarXiv 2403.13002

Mark the obvious cell in the grid and go farthest from it.

Method No. 27Morphological boxclassic

Solve the task as if it belonged to a world far away.

Method No. 05Distant-domain analogyarXiv 2608.03233

Borrow one mechanism from a distant field with the same shape.

Method No. 08Retrieval-grounded inspirationarXiv 2410.14255

Restate the problem three ways; solve the one that changes what winning means.

Method No. 16Analogy-driven reformulationarXiv 2302.12832

Obey one strange instruction and leave no trace of it.

Method No. 28Oblique instructionclassic

Solve three similar problems first, then let them correct the answer.

Method No. 17Thought propagationarXiv 2310.03965

Generate with the rules hidden; bring them back only to choose.

Method No. 14Diverge then convergearXiv 2605.28465

Keep the strongest part of each and build one idea on them.

Method No. 34Candidate fusionarXiv 2608.18931

Breed six ideas, never two on the same mechanism.

Method No. 31Evolutionary populationarXiv 2605.30961

Every rejection becomes a rule for the next round.

Method No. 30Critique as constraintsarXiv 2608.18931

Branch, score, prune the siblings that look alike.

Method No. 33Tree of ideasarXiv 2510.21341

Keep the best idea per cell and aim at the empty ones.

Method No. 32Quality-diversity archivearXiv 2605.09781

27 come from AI research, 5 are classic tools, and 2 are the author’s own.

“Be more creative” is just an adjective. You need a procedure.

method34 is a skill for your AI agent: Claude Code, Codex and others. You give it a task, and it solves it with 3 different methods from the deck below.

1

Give it a task. It sorts the methods into 3 groups by fit and draws 1 at random from each.

In your agent, type:

/method34 <task>
3 ideas, side by side.

Short on time? /method34 light <task> gives 1 idea, the best fit.

2

You get 3 ideas. One method fits, one fits partly, and one is a deliberate mismatch.

3

Draw again, or fuse the best parts into 1 idea.

/method34 again
3 new methods, none repeated.
/method34 fuse
The best parts, in 1 idea.

Run the same prompt on your own machine.

GitHub
Install
Install the method34 skill from https://github.com/gr34tlab/method34.

1. Download the repository into a temporary folder.
2. Copy the skill folder to ~/.claude/skills/method34/. Do not copy the .git folder.
3. Show me the Quick start section of SKILL.md.
Optional: check it first

Want to look inside before you install? Run this one first. It installs nothing.

Check the method34 skill from https://github.com/gr34tlab/method34. Do not install anything.

1. Download the repository into a temporary folder.
2. Read every file, including the Python scripts in scripts/.
3. For each script, tell me these four things:
   - What the script does.
   - If the script reads or writes anything outside the skill folder.
   - If the script makes network calls.
   - If the script asks for keys.
4. Show me the list of files and your findings.

Claude Code · Codex · other agents that work with local files. The check installs nothing.

Free. Run by hand, never automated, never built into a product or a company’s systems. Read the license.

Dmitrii Keshishev, founder of Great Lab, built method34. And this page with it.

Built for ideas: names, campaigns, headlines, visual concepts.

Bring me your brand or product. I put the team together and lead the work. First versions from five working days.

Work with me

I build methods that move AI off its default answer. method34 is the first one made public.

Research with me

Built with Claude Code / method34 / Exa, web search for agents / Ideogram, ChatGPT and Gemini for images / Photoshop / ffmpeg / A human

All sources

AI research · 25 papers

  1. No. 01 Persona setBreaking Homogeneity: Diversifying Persona Sets for Creative LLM OutputsMoon et al., 2026. arXiv 2609.30492
  2. No. 02 Meta-persona anchoringBeyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature ScalingFu et al., 2026. arXiv 2608.02618
  3. No. 03 Paired constraints, No. 13 Scale then dedupeAI and Its Impact on Creativity and Diversity: An Empirical Study of LLM-Generated Product IdeasTerwiesch et al., 2026. arXiv 2607.27553
  4. No. 05 Distant-domain analogyOn the Diversity of Analogy Making in Large Language ModelsShen et al., 2026. arXiv 2608.03233
  5. No. 06 Language of thoughtLanguage of Thought Shapes Output Diversity in Large Language ModelsXu and Zhang, 2026. arXiv 2601.11227
  6. No. 08 Retrieval-grounded inspirationNova: An Iterative Planning and Search Approach to Enhance Novelty and Diversity of LLM Generated IdeasHu et al., 2024. arXiv 2410.14255
  7. No. 10 Forbid the first NIDEAFix: Evaluation Framework for Creative Defixation Prompting in LLMsCarichon et al., 2026. arXiv 2606.00875
  8. No. 11 Verbalized samplingVerbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM DiversityZhang et al., 2025. arXiv 2510.01171
  9. No. 12 Dice-conditioned generationGacha Decoding: Eliciting Diverse Generations Through Instruction FollowingGeng et al., 2026. arXiv 2610.01382
  10. No. 14 Diverge then convergeBeyond One Path: Evaluating and Enhancing Divergent Thinking in Interactive LLM AgentsPark et al., 2026. arXiv 2605.28465
  11. No. 15 Random stimulusAddressing LLM Diversity by Infusing Random ConceptsAgrawal and Goyal, 2026. arXiv 2601.18053
  12. No. 16 Analogy-driven reformulationFluid Transformers and Creative Analogies: Exploring Large Language Models' Capacity for Augmenting Cross-Domain Analogical CreativityDing et al., 2023. arXiv 2302.12832
  13. No. 17 Thought propagationThought Propagation: An Analogical Approach to Complex Reasoning with Large Language ModelsYu et al., 2023. arXiv 2310.03965
  14. No. 18 Conceptual blendingVisual Conceptual Blending with Large-scale Language and Vision ModelsGe and Parikh, 2021. arXiv 2106.14127
  15. No. 19 TRIZ contradictionAutoTRIZ: Automating Engineering Innovation with TRIZ and Large Language ModelsJiang et al., 2024. arXiv 2403.13002
  16. No. 20 Parallel hatsPTFA: An LLM-based Agent that Facilitates Online Consensus Building through Parallel ThinkingGu et al., 2025. arXiv 2503.12499
  17. No. 21 Diverge/converge coachScaffolding Creativity: How Divergent and Convergent LLM Personas Shape Human Machine Creative Problem-SolvingRosenbaum et al., 2025. arXiv 2510.26490
  18. No. 22 Topic-tree distant pairingThe Budget AI Researcher and the Power of RAG ChainsLee and Ma, 2025. arXiv 2506.12317
  19. No. 23 Form constraintLanguage Models are Lazy Poets: Creative Mode Collapse at ScaleMatthews and Hamilton, 2026. arXiv 2610.24223
  20. No. 24 Template rotationDiversity Combining for Multi-Path LLM ReasoningYu et al., 2026. arXiv 2609.38829
  21. No. 29 Wide parallel initialisationInitialization Improves LLM-Driven DiscoverySakarvadia et al., 2026. arXiv 2610.00707
  22. No. 30 Critique as constraints, No. 34 Candidate fusionTest-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the BottleneckRomano et al., 2026. arXiv 2608.18931
  23. No. 31 Evolutionary populationEvoGens: A Population-Based Heuristic Search Framework for Scientific Idea GenerationLi et al., 2026. arXiv 2605.30961
  24. No. 32 Quality-diversity archiveParameter-Efficient Neuroevolution for Diverse LLM Generation: Quality-Diversity Optimization via Prompt Embedding EvolutionGuo et al., 2026. arXiv 2605.09781
  25. No. 33 Tree of ideasMagellan: Guided MCTS for Latent Space Exploration and Novelty GenerationChang, 2025. arXiv 2510.21341

Classic tools · 5

  1. No. 04 Reframe and name the frameWinston, MIT 6.034
  2. No. 09 Seen-set exclusionWinston, non-redundant generator
  3. No. 26 Worst possible ideaClassic tool
  4. No. 27 Morphological boxClassic tool
  5. No. 28 Oblique instructionClassic tool

The author’s own · 2

  1. No. 07 Maker + knower pairThe author’s own method
  2. No. 25 Assumption ledgerThe author’s own method