Adversarial Planning Council

A pre-build gate in which a council of persona sub-agents argues against your idea before any code is written, and only a surviving idea proceeds to the build. You spin up several agents with distinct, deliberately opposed roles — one makes the strongest case for the idea, one attacks it and looks for every hole, one gathers evidence from a purely objective/research angle — run them over the plan or idea (not the output), require every point to be "backed by something real, not just vibes," and collapse their arguments into one verdict: Go / Reshape / Kill. The council exists to counter a specific, standing failure mode: LLMs lean sycophantic — arrive excited and the model wants to make you more excited and start building, even on a holed idea — so agreement from a single assistant is uninformative. Manufacturing structured disagreement, with evidence, is how you get a signal instead of a hype-echo. In the source it is a named, packaged ritual ("the roast" — "a file you drop into Claude" that runs the whole council automatically), and, crucially, its persona members share the operator's standing context (what's working in the business, current priorities) so the critique is grounded in reality rather than generic.

This is distinct from Advisor Mode (escalation to a smarter model when an executor gets stuck — a capability upgrade, not an oppositional panel) and from a plain multi-agent fan-out: the defining move is engineered adversarial diversity over one artifact — the idea — culminating in a build/don't-build decision. It is the plan-stage sibling of Pre-Deployment Validation: that gate checks the premises of a build before you pay to run; the council checks the soundness of the idea before you pay to build.

Claims

Because models lean sycophantic, agreement from a single assistant is not evidence an idea is good; you must engineer disagreement to get signal. principle — durable: a people-pleasing model amplifies the operator's own excitement, so the informative move is to force the strongest case against the idea, not to ask "is this good?" and be reassured.
Run the critique on the idea/plan before the build, and let only a surviving (or reshaped) idea proceed. best practice — context: work where building the wrong thing is expensive relative to the cost of the critique; on throwaway or cheap-to-try ideas, just build and learn (Agentic Simplicity). The gate earns its keep when the plan is load-bearing — the Spec-Driven Development "front-load attention onto planning" discipline, staffed by adversaries.
Give the council opposed personas — advocate, critic, objective evidence-gatherer — rather than N identical reviewers, so it surfaces failure modes redundancy would miss. best practice — context: pre-build idea review; role diversity is what makes it a red-team and not theater. Mirrors the "distinct lenses beat identical refuters" discipline used in adversarial verification.
Require every point to be backed by something real, not vibes, and end in one clean verdict: Go / Reshape / Kill. best practice — context: decision-forcing reviews; the evidence requirement and the single verdict are what convert a sprawling critique into an actionable gate. This is Evidence-Gated Completion's "prove it" discipline applied to the plan rather than the output.
Feed the council the operator's standing context so its critique is grounded in the real situation, not generic objections. observation — the source notes each council member "knows what's currently working in my business," which is what makes the roast better than a cold review; it is Context Substrate / Self-Improving System memory reused at the planning gate.
The value is produced by the peer-review round, not by the parallel answers — the cross-review is what surfaces what every advisor missed. principle — durable, and a sharpening added by a second, independent instance of the pattern (llm-council-skill): five advisors answer in parallel, then review each other, and the artifact's payoff section is literally headed "Blind Spots the Council Caught… what every advisor missed" — in the worked example, unquantified unit economics and a fulfillment-capacity ceiling, neither of which was any single persona's answer. Persona count buys coverage; the review pass buys the blind spots. If you cut a stage, cut a persona, not the cross-review.
The council generalises past the pre-build gate to any consequential decision. observation — the second instance runs it on an operating question ("chase new subscribers or raise prices on the loyal base?") rather than a build/don't-build, and returns a verdict plus blind spots rather than Go / Reshape / Kill. The shape survives; only the verdict vocabulary changes.
Persona diversity can substitute for model diversity. observation — Karpathy's original method is described as polling different LLMs (ChatGPT, Gemini, Claude, Perplexity) against one another; llm-council-skill runs all five advisors as sub-agents on a single model and its author reports it "still works quite well." Asserted from use, not measured — a testable refinement, not a settled result, and a live question for anyone paying for multi-vendor fan-out.
Three independent sources now converge on this pattern, packaged and shipped as a drop-in file: a personal "roast" ritual; the public tenfoldmarc/llm-council-skill (advisors: contrarian, first-principles thinker, expansionist, outsider, executor); and a STORM research skill (lenses: practitioner, academic, skeptic, economist, historian). observation — different authors, different framings, same primitive.
The council generalises past decision gates to research synthesis: the same parallel-personas + cross-review shape produces a verified multi-perspective briefing, not just a Go/Reshape/Kill. observation — the STORM instance runs the panel to build knowledge rather than to kill an idea: five lenses in parallel → a "contradiction map" (where they disagree, whose evidence is strong/weak) → synthesis → an adversarial peer-review pass that verifies every citation against its primary source. The intent shifts from screening to sourcing; the primitive is unchanged, and the cross-review round is again where the value lands (it names the assumption the briefing rests on and the missing lens none of the personas occupied).
An author-constructed panel of personas on one model with a shared framing yields correlated, not independent, judgments — so agreement across lenses is a strong hypothesis, not independent proof. principle — durable, and the sharpest available answer to this page's open "persona-diversity substitutes for model-diversity" question: the STORM artifact prints the caveat on its own face ("the panel was author-constructed… where they agree, treat it as a strong hypothesis, not independent proof"). The consequence is that trust must ride on the verification pass, not on the panel agreeing with itself — which is why the research variant bolts an evidence-gated citation check onto the council (Evidence-Gated Completion).
A fourth source takes the model-diversity horn of this page's open question explicitly, and names the mechanism: the echo chamber is a property of the model's lineage, not of its prompt. observation — grill-me-codex runs a two-member council (Claude Fable 5 vs OpenAI Codex) rather than N personas on one model, and its README states the reason on its face: "the same model that plans the build and writes the build can't be trusted to grade its own work — it's an echo chamber. A different provider catches what Claude structurally can't see in itself." This does not settle persona-vs-model diversity — no measurement is offered on either side — but it converts this page's own correlated-not-independent caveat into a named counter-position from a practitioner who shipped on it. The axis now has its own page: Cross-Model Independence.
If uncorrelated errors are what a roster actually buys, model diversity may buy with two members what persona diversity needs five to approximate. observation — the cross-model instance runs a council of two and reports convergence in 2 rounds / 8m 14s; the persona instances run five. Asserted from the shape of the patterns, not measured. See Bounded Fan-Out.
The verdict vocabulary flexes again — a cross-model council terminates on converge-or-cap, not Go/Reshape/Kill. observation — two independent critics have no guarantee of agreeing, so the cross-model instance bounds the argument at 5 rounds and defines a fallback rather than forcing a verdict. The primitive survives; how it is made to end is a design choice. See Bounded Negotiation with Fallback.
A single dissenter is enough to break a consensus — the roster may not need to be large, it needs to be non-empty. (observation — as recounted by a fifth source) — in the Asch line-matching experiment Sandeep Swadia recounts, three of four participants gave a knowingly wrong answer to match seven confident actors, and when just one actor began answering correctly, most participants recovered the right answer too. The asymmetry is the operationally interesting half: manufacturing conformity took a crowd, breaking it took one voice. If it transfers, it bears directly on Bounded Fan-Out and on this page's persona-count question — the council's value may come from the existence of an opposed member rather than from the breadth of the roster. Recounted second-hand from a human-psychology experiment, with no claim that it transfers to model panels; a suggestive analogy, not a result.
The devil's-advocate move is prescribed from outside engineering as a general-purpose thinking tool, and explicitly as a fallback to a human dissenter. observation — the same source: "when everyone around you agrees, make AI your voice of dissent, your devil's advocate. Tell it to give you an opposite argument. Make the strongest case that this consensus is wrong." Worth noting his ordering — the primary advice is go find the person who disagrees, and the model is what you reach for when no such person is available. The vault's practitioner sources reach for the model first; a non-practitioner ranks a human dissenter above it, which is a quiet challenge to how readily these councils substitute for real disagreement.

Advisor Mode — the contrast case: advisor mode escalates to a smarter model when stuck (a capability pull); the council manufactures opposition among peers to test an idea. Different axis.
Pre-Deployment Validation — the sibling gate one step later: it validates the premises of a build before you pay to run; the council validates the idea itself before you pay to build.
Spec-Driven Development — "plan-first, front-load attention on the plan" is where the council operates; it is the adversarial staffing of the planning step.
Evidence-Gated Completion — the same "backed by something real, prove it" gate, applied to the plan instead of the finished output.
Negative Prompting — the council is an operator-set constraint that forces the model to argue against, rather than please; both counter default sycophancy.
Agent Supervision — the human still owns the Go/Reshape/Kill call; the council structures the evidence the operator decides on.
Workflows vs Agents — the council is a fixed, repeatable workflow of role-typed sub-agents, not an open-ended agent, which is what lets it ship as a drop-in file.
Context Substrate — the standing memory the council members read so their critique is grounded.
Role-Typed Agent Roster — the same persona primitive with the opposite intent: a roster's personas divide labour, a council's are engineered to disagree.
Public Skill Adoption — how the pattern actually spreads: as a two-file repo you clone.
Agent Communication Topology — the council's cross-review can be a mesh debate or a hub-run review over star-topology outputs; STORM takes the second route (subagents that never talk, plus a contradiction+peer-review stage). Topology is not what makes it a council; the review round is.
Bounded Fan-Out — a council is a fixed small roster of opposed personas; the bound is what lets it ship as an always-the-same-panel drop-in.
Cross-Model Independence — the axis this page's open question was really about: the council manufactures opposition (the fix for sycophancy); that page argues opposition sourced from one model is correlated, and that independence must be bought across provider lineages.
Bounded Negotiation with Fallback — how a council of genuinely independent members is made to terminate: a round cap plus a defined fallback, instead of a forced verdict.
Distillate: How Claude Is Creating a New Generation of Millionaires
Distillate: 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)
Distillate: STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill — a third instance (STORM), turned to verified research synthesis, whose artifact admits its author-constructed panel is correlated-not-independent.
Falsification-First Questioning — the primitive beneath this page: the council is "ask the question that could return a bad answer," staffed and automated.
Distillate: Fable 5 + GPT 5.6 Sol = CHEAT CODE — a fourth instance that takes the other horn: a two-member cross-vendor council (Fable 5 vs Codex), justified by an explicit echo-chamber argument, and terminated by a round cap rather than a verdict.
Distillate: This Skill Makes You Dangerous In The AI Era — a fifth, non-practitioner instance: AI as devil's advocate, ranked below finding a human who disagrees, plus the one-dissenter-suffices finding.
Stakeholder Clone — the near-neighbour that takes the opposite horn on both axes: it clones a specific named real person (not an engineered adversary) to react to output before submission (not an idea before the build). Same persona-sub-agent primitive, fidelity instead of opposition.
Distillate: You're the Problem, Not Claude — Six Fixes to 10x Output — introduces Stakeholder Clone (ask-the-board / clone-your-manager / internal-focus-group), the fidelity-to-a-real-person variant of the persona-panel primitive.

Adversarial Planning Council

Claims

Related

Linked from