Convene history's greatest minds to think through your hardest problems.
A single-perspective prompt produces a single-perspective answer.
Ask an LLM "what should I do about X?" and you get one voice, one frame, one recommendation. It sounds authoritative. It is also narrow in exactly the way you cannot see, because the narrowness matches the framing of your question.
Twenty named minds produce genuine surprise.
When you ask Claude, GPT-4 or Gemini to convene Daniel Kahneman, Seneca, Simone de Beauvoir, Nassim Taleb and sixteen others, each speaking from their actual intellectual framework, something different happens. The model activates deep knowledge of how each thinker actually thinks, not a caricature but a substantive simulation. Kahneman does not just say "watch out for biases"; he names the specific bias operating in your situation and explains the experimental evidence behind it. Seneca does not just say "memento mori"; he reframes your timeline problem as a problem of attention misallocation. And when Kahneman and Seneca agree on something from completely different premises, you have a signal that no single-perspective answer could have produced.
The council pattern is prompt engineering as intellectual architecture. It works because large language models have ingested the complete published works, interviews, lectures and secondary literature of thousands of thinkers. The named mind is the retrieval key. The council structure is the forcing function that makes the model use all of them instead of collapsing to the median.
-
Name the minds. Not "an economist" but "Daniel Kahneman." Not "a philosopher" but "Simone de Beauvoir." The name activates the model's deep knowledge of their actual thinking patterns, not a generic disciplinary perspective.
-
Define their lenses. Each mind has a primary cognitive function: critic, builder, ethicist, historian, futurist, systems thinker, practitioner, contrarian, synthesizer, domain specialist, outsider, advocate. The function tells the model how each mind should engage with the problem, not just what they know.
-
Cross-pollinate. After each mind speaks, the council identifies unexpected agreements, genuine disagreements, and what emerges from the intersections. This is where the surprise lives. When a Stoic emperor and a behavioural economist reach the same conclusion through different reasoning, you have found something load-bearing.
-
Synthesize. What the council says in unison. Where it splits. What it cannot know. And one actionable recommendation grounded in the full read.
The difference between "an economist" and "Daniel Kahneman" is the difference between a Wikipedia summary and a mind.
When you name Kahneman, the model does not produce generic economic reasoning. It produces prospect theory, loss aversion, the planning fallacy, System 1 vs System 2, the specific experimental designs Kahneman used to demonstrate each phenomenon. It produces the way Kahneman thinks, which is different from the way Amartya Sen thinks, which is different from the way Charlie Munger thinks, even when all three are reasoning about the same decision.
Named minds also constrain hallucination. The model knows what Kahneman would and would not say. It knows Seneca would not invoke regression analysis, and Dijkstra would not appeal to user experience research. The name is both a retrieval key and a guardrail.
Twenty minds is not arbitrary. It is the minimum count that reliably produces:
- At least three disciplines represented
- At least two centuries of thought
- At least one genuine contrarian
- At least one practitioner who asks "does this work on Monday morning?"
- Enough diversity that the synthesis cannot collapse to a single frame
- Pick a pre-built council below that matches your problem domain.
- Copy the prompt from
prompt_templates/claude.md(orgpt4.mdorgemini.md). - Replace the
[PROBLEM]placeholder with your specific question. - Paste into your preferred LLM.
- Read the synthesis. The surprise is usually in the cross-pollination section, where minds you did not expect to agree converge on something you had not considered.
| Council | Domain | Link |
|---|---|---|
| Career Strategy | Career transitions, job decisions, professional identity | councils/career_strategy.md |
| Technology Strategy | Architecture decisions, build vs buy, technology bets | councils/technology_strategy.md |
| Policy Analysis | Public policy, institutional design, governance | councils/policy_analysis.md |
| Creative Writing | Craft, voice, structure, publishing | councils/creative_writing.md |
| Startup Strategy | Founding, scaling, fundraising, product-market fit | councils/startup_strategy.md |
See templates/custom_council.md for the step-by-step protocol, or read the full specification in protocol/council_protocol.md.
The short version:
- Define your problem with precision. Not "help me with my career" but "I am 37, employed below my capability, with two children and a contract ending in eight weeks. Given these constraints, what is the optimal move?"
- Select twenty minds across at least three disciplines and two centuries.
- Assign each mind a primary function from the twelve functions.
- Run the protocol (first round, cross-pollination, synthesis, practical consequence).
Every council member serves one primary function. The functions ensure coverage: a council of twenty historians would produce depth without breadth, twenty futurists would produce speculation without grounding. The twelve functions are:
| Function | Role | Example minds |
|---|---|---|
| Critic | Tears down weak arguments before they travel | Kahneman, Popper, Hitchens |
| Builder | Constructs solutions from constraints | Drucker, Engelbart, Wozniak |
| Ethicist | Names what is right and what it costs | Rawls, Nussbaum, Frankl |
| Historian | Maps the present onto past patterns | Arendt, Braudel, Du Bois |
| Futurist | Projects consequences forward ten years | Kay, Thiel, Engelbart |
| Systems Thinker | Sees feedback loops and second-order effects | Meadows, Ostrom, Taleb |
| Practitioner | Asks "yes but does it work on Monday morning?" | Graham, Blank, Horowitz |
| Contrarian | Holds the opposite position to stress-test | Taleb, Thiel, Munger |
| Synthesizer | Finds the thread across disciplines | Munger, Sen, Morrison |
| Domain Specialist | Brings deep technical knowledge in the problem's field | Varies by council |
| Outsider | Sees what insiders cannot because they are too close | Varies by council |
| Advocate | Represents the interests of those affected but not present | Farmer, Fanon, Butler |
Full descriptions with worked examples: protocol/twelve_functions.md
twenty-mind-council/
README.md # This file
LICENSE # MIT
.gitignore
protocol/
council_protocol.md # The full seven-step protocol
twelve_functions.md # The twelve cognitive functions in detail
councils/
career_strategy.md # Pre-built: career decisions
technology_strategy.md # Pre-built: technology decisions
policy_analysis.md # Pre-built: policy and governance
creative_writing.md # Pre-built: writing craft
startup_strategy.md # Pre-built: founding and scaling
templates/
custom_council.md # Template for building your own
prompt_templates/
claude.md # Ready-to-paste prompt for Claude
gpt4.md # Ready-to-paste prompt for GPT-4
gemini.md # Ready-to-paste prompt for Gemini
Here is a compressed demonstration. The full worked example is in councils/career_strategy.md.
Problem: "I am 38, underutilized in a mid-level role, with a family. A former colleague offers me a senior position at a younger, riskier organization. Should I take it?"
What happens when twenty minds engage:
Kahneman identifies three biases operating in the framing (loss aversion, status quo bias, planning fallacy). Seneca reframes the timeline: "the question is not whether you can afford the risk of change but whether you can afford the certainty of staying." Taleb rejects the binary entirely and proposes a barbell strategy. Frankl asks whether the current role allows meaning. Ibarra says stop deliberating and start testing possible selves through small experiments. Du Bois asks whether the "stable" role is stable by design or stable because the institution has capped you. Brene Brown insists the partner must be part of the conversation.
The cross-pollination surprise: Seneca (1st century Stoic) and Kahneman (21st century behavioural economist) converge on the same insight from completely independent frameworks: the feeling of safety in the current role is not evidence of safety; it is evidence of status quo bias. When two thinkers separated by two millennia reach the same conclusion through different reasoning, you have found something load-bearing.
The practical consequence: Accept the new role, but only after three actions. First, have the vulnerable conversation with your partner (Brown). Second, build a six-month financial buffer (Buffett, Hoffman). Third, begin one side project that compounds regardless of employer (Taleb's barbell). The reasoning is that seventeen of twenty frameworks converge on the same directional conclusion, and the risk is managed by the buffer, the side project, and Hoffman's Plan B.
This is what twenty minds produce that a single-perspective answer cannot: genuine surprise, grounded in the specific frameworks of thinkers whose work the model knows deeply, with a synthesis that names what is known, what is disputed, and what only the decision-maker can determine.
It works, and the reason is structural. Large language models have ingested the complete published works of these thinkers. When you name "Daniel Kahneman," the model does not generate random economic reasoning; it activates the specific concepts, experimental evidence, and argumentative patterns that Kahneman used across his career. The council format forces the model to use all twenty of these knowledge clusters simultaneously, which produces outputs that are meaningfully different from what any single prompt would produce. The surprise, the unexpected agreement between Seneca and Kahneman, is not fabricated; it is discovered by the model through the multi-perspective structure.
Twenty is the empirical minimum for reliable coverage. At ten, you can cover three disciplines but not the full function set (critic, builder, ethicist, etc.). At fifty, the contributions become too thin and the cross-pollination becomes noise. Twenty minds with twelve cognitive functions produces genuine intellectual diversity without dilution. This is a practical finding from running hundreds of councils, not a theoretical constraint.
Yes. A twelve-mind council (one per function) works for many problems. A five-mind council works for narrow technical questions. But the cross-pollination step, which is where the distinctive value lives, requires at least ten minds to produce meaningful unexpected agreements.
Absolutely. Take Kahneman from the career council, Dijkstra from the technology council, and Fanon from the policy council. The protocol works with any combination of twenty named minds. The custom council template in templates/custom_council.md walks you through the selection process.
All three major LLMs (Claude, GPT-4, Gemini) handle the council pattern well, with different strengths. Claude produces deeper per-mind contributions and stronger cross-pollination. GPT-4 produces more concrete practical consequences. Gemini handles the full twenty-mind protocol in a single response more reliably due to its larger context window. Ready-to-paste prompts for all three are in the prompt_templates/ folder.
Yes, with one caveat: the council will not make your decision for you. It will clarify what is at stake, identify the biases in your thinking, and surface perspectives you had not considered. But personal decisions involve values that only you hold, and the council's synthesis always includes a "what the council cannot know" section that names these irreducible personal elements.
Each mind's 2-3 sentence contribution in the first round is not an answer; it is a lens. The value of the first round is to produce twenty different framings of the same problem. Read them for the framings, not for the recommendations.
This is where the council produces insights that no single-perspective prompt can match. Pay particular attention to:
-
Unexpected agreements across centuries. When Seneca (1st century) and Kahneman (21st century) converge on the same insight through different reasoning, you have found something that is robust across time and methodology. These convergences are the highest-confidence findings.
-
Unexpected agreements across disciplines. When a philosopher and an economist reach the same conclusion, the conclusion is probably not discipline-dependent. This matters because discipline-dependent conclusions may be artefacts of the analytical framework rather than features of reality.
-
The disagreement that will not resolve. Some disagreements in the council represent genuine trade-offs that cannot be optimized away. Taleb's "diversify for survival" and Thiel's "concentrate for monopoly" are both correct, under different conditions. The council's job is to name the trade-off, not to resolve it. Your job is to decide which condition you are in.
Every council synthesis includes a section on what the council cannot determine: the personal values, the private circumstances, the relationship dynamics, the emotional reality that only the decision-maker holds. This section is not a disclaimer; it is the boundary between analysis and judgment. The council provides the analysis. The judgment is yours.
- When the problem changes. New information (an offer arrives, a deadline moves, a constraint changes) makes the original council outdated. Re-run with the updated problem.
- When you disagree with the synthesis. If your gut says the council is wrong, that is valuable data. Re-run with a different council composition that includes minds representing your position, and see if the synthesis holds or shifts.
- When the council splits evenly. A council that splits 10-10 on a question has not produced a recommendation; it has identified a genuine dilemma. Run a second, narrower council on the specific point of disagreement.
Large language models store knowledge in distributed representations. When you prompt with "Daniel Kahneman," the model activates a cluster of associations: prospect theory, loss aversion, Thinking Fast and Slow, the Nobel Prize work with Amos Tversky, the heuristics and biases research programme, the specific experimental designs, the argumentative style (empirical, cautious, evidence-grounded). This cluster is qualitatively different from the cluster activated by "an economist" or even "a behavioural economist."
The named mind is a high-fidelity retrieval key. It activates not just the thinker's conclusions but their methodology, their style of reasoning, and their known positions on adjacent topics. This is why the council produces genuinely different contributions from each mind, rather than twenty variations on the same generic analysis.
A single-perspective prompt activates one knowledge cluster and produces one analysis. Twenty named minds activate twenty knowledge clusters simultaneously. The cross-pollination step forces the model to identify connections between these clusters that it would not produce from any single cluster alone.
The surprise, the unexpected agreement between Seneca and Kahneman, is not fabricated. It is discovered by the model through the multi-perspective structure. The model has deep knowledge of both Seneca and Kahneman, but in a single-perspective prompt, it would only activate one of them. The council structure activates both and asks "where do they converge?" The answer is often genuinely surprising because the connection crosses disciplinary boundaries that the model (and the user) would not normally bridge.
Without function assignments, twenty minds tend to produce twenty analyses of similar type: each mind states their view, but they all do the same kind of intellectual work (typically analysis or recommendation). The function assignments force diversity of cognitive operation: the Critic tears down, the Builder constructs, the Ethicist names moral costs, the Practitioner checks implementation. This diversity of operation is what produces the rich interaction patterns in cross-pollination.
This pattern emerged from practice, not from theory. Over months of working with LLMs on problems that required genuine multi-perspective analysis, a repeatable structure crystallized: name the minds, define their lenses, force cross-pollination, synthesize honestly. The protocol was refined through hundreds of council sessions across career strategy, institutional analysis, editorial decisions and policy questions.
The key discovery was that the named mind is a fundamentally different retrieval mechanism than the role description. "An economist" gives you the median of economics. "Daniel Kahneman" gives you prospect theory, loss aversion, and the specific experimental evidence that supports each claim. The name is a precision instrument.
The five pre-built councils cover career, technology, policy, creative writing, and startups. Here are starter lists for councils in other domains, ready for anyone to build out with framework summaries and worked examples:
Atul Gawande (checklists, complexity), Paul Farmer (structural violence, accompaniment), Siddhartha Mukherjee (the emperor of all maladies, biography of disease), Abraham Verghese (the importance of the physical exam, bedside manner), Oliver Sacks (the narrative of neurological conditions), W. Edwards Deming (quality management), Florence Nightingale (data-driven reform), Don Berwick (Triple Aim), Christine Cassel (geriatrics, end-of-life care), Ira Byock (dying well), Lisa Sanders (diagnosis as detective work), Howard Brody (the physician's covenant), Ivan Illich (medical nemesis), Aaron Antonovsky (salutogenesis), Gabor Mate (trauma and addiction), Daniel Kahneman (diagnostic bias), Marcia Angell (the truth about drug companies), Christina Maslach (clinician burnout), Martha Nussbaum (dignity in care), Amartya Sen (health as capability).
John Dewey, Paulo Freire, Maria Montessori, E.D. Hirsch, Howard Gardner, Carol Dweck, Angela Duckworth, Sugata Mitra, Ken Robinson, Ivan Illich, Lev Vygotsky, Jean Piaget, bell hooks, Diane Ravitch, Dylan Wiliam, James Baldwin, Hannah Arendt, Seymour Papert, Linda Darling-Hammond, Yong Zhao.
Warren Buffett, Charlie Munger, Benjamin Graham, John Bogle, Nassim Taleb, Daniel Kahneman, Seneca, Ramit Sethi, Morgan Housel, JL Collins, Vicki Robin, Peter Lynch, Howard Marks, Ray Dalio, Suze Orman, Thomas Piketty, Amartya Sen, Hernando de Soto, Muhammad Yunus, Adam Smith.
Esther Perel, John Gottman, Alain de Botton, bell hooks, Brene Brown, Simone de Beauvoir, Erik Erikson, Viktor Frankl, Carl Jung, Irvin Yalom, Donald Winnicott, Mary Ainsworth, Harriet Lerner, Terry Real, Sue Johnson, Gabor Mate, James Baldwin, Audre Lorde, Fred Rogers, Thich Nhat Hanh.
Rachel Carson, James Lovelock, Donella Meadows, Elinor Ostrom, Wangari Maathai, Amory Lovins, Bill McKibben, Naomi Klein, Paul Hawken, Vandana Shiva, Robin Wall Kimmerer, E.O. Wilson, Herman Daly, Tim Jackson, Kate Raworth, Hans Rosling, Johan Rockstrom, Katharine Hayhoe, Michael Mann, Sunita Narain.
Contributions welcome. To add a new pre-built council:
- Follow the format in
councils/career_strategy.md(table of minds, framework summaries, worked example with first round, cross-pollination, synthesis, and practical consequence). - Ensure the council meets the selection criteria (three disciplines, two centuries, one contrarian, one practitioner, one advocate).
- Include framework summaries for each mind that go beyond Wikipedia-level descriptions. The framework summary should name specific concepts, books, papers, and argumentative patterns.
- The worked example should demonstrate genuine cross-pollination: unexpected agreements, genuine disagreements, and emergent insights.
The council pattern has been applied to problems including:
Career transitions. A professional in hospitality with credentials in international programme coordination used the career strategy council to evaluate five simultaneous pivot options across three countries. The council's cross-pollination revealed that the real question was not "which job?" but "which position lets me become the person my children need to see?", a reframe that no single-perspective analysis could have produced. Erikson's generativity stage, Du Bois's double consciousness, and Ibarra's working identity converged on an identity-construction frame that was more useful than any job-comparison matrix.
Technology architecture. A team debating microservices vs monolith used the technology strategy council. The council's sharpest finding was not the architecture recommendation but Brooks's diagnostic: "Does one person understand this entire system?" The answer was no, which meant the architecture question was premature. The team spent three months on documentation and specification (Lamport) before making the architecture decision, and by then the decision was obvious.
Policy design. A development organization debating programme design for rural education used the policy analysis council. Scott's metis framework (what local knowledge will this programme destroy?) combined with Duflo's empirical rigour (what does the evidence from comparable interventions show?) produced a design that neither perspective alone could have generated: a programme that preserved local teaching practices while adding measurable learning outcomes.
Personal decisions. A parent deciding whether to relocate for a career opportunity used the career strategy council with an emphasis on Brown's Advocate function. The council's most important finding was in the "what the council cannot know" section: the children's own preferences. The parent had not asked them. After asking, the decision became clear, and it was different from what the council recommended.
Writing and publishing. A writer with a completed manuscript used the creative writing council to diagnose structural problems. Morrison's principle ("find the sentence the novel exists to deliver") unlocked a revision strategy that three years of workshop feedback had failed to produce.
These applications demonstrate the council's core value proposition: it does not produce answers that a smart person could not eventually reach alone. It produces those answers in ninety minutes instead of ninety days, and it surfaces perspectives that the decision-maker's own cognitive habits would have suppressed.
The council pattern is not magic. It has real limitations:
-
The model's knowledge has a training cutoff. Minds who published after the cutoff will have thinner representations. Living thinkers whose most recent work is post-cutoff will be simulated based on older work.
-
The model does not actually contain the minds. It contains a statistical representation of their published work. The simulation is good enough to produce genuine intellectual surprise in cross-pollination, but it is not a substitute for reading the original works.
-
The council cannot know what you have not told it. If your problem statement omits key constraints, the council will produce recommendations that do not fit your real situation. The quality of the output is directly proportional to the honesty and specificity of the input.
-
The council does not replace domain expertise. For a medical decision, consult a doctor. For a legal decision, consult a lawyer. The council provides multi-perspective analysis; it does not provide professional advice.
-
The council can produce false confidence. When seventeen of twenty minds agree, the recommendation feels authoritative. But the agreement may reflect the model's training distribution (most texts agree on this point) rather than genuine intellectual convergence. Use the council as one input among many, not as the oracle.
David Dabert
MIT. See LICENSE.