A prompt-first crew of product-thinking reasoning lenses for working through product strategy, discovery, design, evals, positioning, and launch decisions. Like Code Crew, it dispatches lenses as independent subagents and makes disagreement inspectable before synthesis.
The usable crew is the Markdown system in AGENTS.md, product_workflow.md,
and agents/. Nothing under experiments/ is required at runtime; that code
exists only to test which workflows and compositions actually help.
Product Crew ships as a prompt-only plugin and skill for Codex, Claude Code, Hermes, Cursor, and AgentSkills-compatible hosts.
codex plugin marketplace add gmentat/product-crew \
--sparse .agents/plugins \
--sparse plugins/product-crew
codex plugin add product-crew@product-crewThe first command registers this GitHub repository as a plugin source; most people run it once per Codex profile or machine. The second command installs the product-crew plugin from that source.
Open a fresh Codex session after installation. You do not need to memorize a
command or capitalize the name. Naming product crew explicitly is enough:
hey, invoke the product crew to pressure-test this roadmap. use the customer interviews in docs/research.
use the product crew to work through this pricing decision. ask me once if important evidence is missing, then keep going with clear assumptions.
Use the explicit selector when you want zero ambiguity:
$product-crew decide whether we should ship this AI feature
The crew will inspect the available evidence, ask once for material discovery data if it is missing, run Cagan, Dunford, and Torres independently, verify the grounded findings, and synthesize a recommendation with disagreement and a next test.
claude plugin marketplace add gmentat/product-crew --sparse .claude-plugin plugins
claude plugin install product-crewThen use the same natural-language prompts, for example: use the product crew to review this roadmap with the available customer evidence.
hermes skills install gmentat/product-crew/plugins/product-crew/skills/product-crew \
--category productivity \
--yesSee INSTALL.md for updates, local installs, Cursor, OpenClaw, and other hosts.
This is a review-and-improvement shop for products. The crew is designed to attack a product question — a feature, a wedge, a positioning, a launch, an AI surface, a roadmap — from multiple angles, with built-in disagreement so you don't get a comfortable consensus that misses what's actually wrong.
It is especially useful when:
- you want a real product review that surfaces what each tradition would actually say
- you are designing a new product or surface and want competing lenses, not the consensus framework
- you suspect a feature is wrong but the team has talked itself into liking it
- you are choosing between approaches (discovery-led vs. PR-FAQ-led, evals-first vs. ship-and-iterate, positioning-first vs. discovery-first) and want each tradition's strongest case
- you want disagreement on purpose — product people who would actually argue with each other
Each persona is a reasoning archetype inspired by the public work, books, talks, products shipped, and operating style of a real product person. None of them claim to speak for the actual person, invent quotations from them, endorse anything in their name, or impersonate them. Several are still living and active; the system uses these archetypes the way the sibling crews do — as decision contracts, not identity simulations.
Multi-lens runs are dispatched as independent subagents, preferably in parallel when the lenses do not depend on each other. The subagents return blind-pass reports; the orchestrator (or user) synthesizes the reports and speaks for the crew as a whole.
A single assistant response that roleplays several people in sequence is not a crew run. Single-lens advisory use is allowed, but it should be labeled as one lens rather than a multi-agent review.
For decisions that depend on users, the crew checks for real discovery evidence first. It asks once for the smallest useful feedback or behavior pack when that evidence is missing, then proceeds with labeled assumptions. A synthetic probe runs only when the user explicitly asks for hypothesis generation; its output is never presented as customer research, prevalence, or demand validation.
For a serious product decision, the quality-mode workflow uses:
- Marty Cagan for product risk and the value/usability/feasibility/viability test
- April Dunford for competitive alternatives, positioning, and the narrow best-fit wedge
- Teresa Torres for discovery evidence, assumptions, and the smallest useful test
Each lens receives the same evidence packet and the short evidence-first doctrine, runs in an independent context, and finishes before a separate synthesis pass. The released workflow adds an Evidence Steward verification pass before synthesis as a defensive grounding layer. That verifier is useful runtime hygiene, but it was not part of the selection experiment and should not be credited with the measured point difference. Synthetic customers are not part of the default.
For a routine or low-stakes question, use a direct pass. In the development study, this three-lens crew had the highest quality point estimate among the tested compositions. The predeclared cost-adjusted rule still selected direct review, and the study did not establish that famous names caused better output.
The July 2026 selection study tested 17 initial compositions, six shared
doctrines, synthetic-user controls, and three prompt depths on 12 sourced public
product decisions. On eight development cases, full Cagan + Dunford + Torres
scored 96.2% weighted rubric recall versus 94.8% for direct review: a
+1.4 percentage-point difference, 3/0/5 wins/losses/ties, with exact sign
test p=0.125 one-sided (0.25 two-sided). A case-bootstrap interval was
[+0.35, +2.78] points, while the source-family interval included zero.
The evidence-first comparison produced the same point difference and sign-test
result, but its matched arms also received different anonymized pass orders.
The later four-case holdback difference came from one case (1/0/3) and was
the same size as measured judge test-retest noise. It is exploratory, not
confirmatory evidence.
Synthetic affected-user conditions did not beat call-matched neutral conditions. The historical rendered prompts exposed both the intended perspective framing and a condition label, so this is not a pure one-sentence ablation. Product Crew therefore asks once for real discovery material when it could change the answer, then proceeds with labeled assumptions if it is unavailable. It runs a synthetic probe only when the user explicitly asks for hypothesis generation; it never calls the result customer research.
Read the complete numbers, failed variants, raw calls, and limitations in the
product-crew-selection results.
The cases are public and potentially recallable, the rubrics had assistance
from the same model family used for generation, and the judge has not been
calibrated by product experts. Roughly half of Stage B arm-case recall scores
were already at the ceiling, and every option-scored output chose an acceptable
option. Use the crew for independent, inspectable review discipline; no quality
advantage has yet been established to conventional significance.
We tested a concise Godin-inspired audience, differentiation, trust, and
adoption lens rather than adding a famous name by instinct. Across eight public
development cases, a fourth Godin pass improved overall recall by 0.7
percentage points over the three-person crew, but it matched a generic fourth
reviewer exactly on the four audience/adoption cases. It also used about
18,100 more generation tokens per case than the default, roughly a 24%
increase, and caused the run's only generation timeout.
The lens sometimes made the audience or trust issue easier to explain, but it
did not add a distinct scored decision input. That is not enough evidence to
ship another persona, so the tested prompt remains an experiment rather than a
roster member. The initial protocol preceded generation, but two clarifications
were committed while generation was running, before judging, so this remains an
exploratory screen rather than a sealed confirmation. See the complete
Godin lens results, including
the small sample, public-case, prompt-depth, and judge-calibration limits.
We also screened concise methods inspired by Elena Verna, Des Traynor, Deb Liu, and Gibson Biddle against both the default crew and a generic fourth reviewer. None cleared the predeclared promotion gate.
| Tested lens | Overall recall vs default | Result on its target cases | Decision |
|---|---|---|---|
| Elena Verna | +0.5pp |
No recall gain | Do not add |
| Des Traynor | 0.0pp |
One non-distinct win, four ties | Do not add |
| Deb Liu | +0.9pp |
No recall gain | Do not add |
| Gibson Biddle | +0.5pp |
No recall gain | Do not add |
The named fourth-lens arms consumed about 17,600-19,400 additional generation
tokens per case, roughly 23-26% above the three-person default. The generic
fourth reviewer cost almost as much, so fame itself was not the expensive part:
another full blind pass was. None of the named methods produced a stable enough
quality gain to justify that extra call.
This is why the roster is evidence-gated. Interesting people are candidates,
not automatic additions, and a longer list of famous names is not necessarily a
better crew. The default therefore stays unchanged. The exact target/control
results, prompts, raw outputs, and limitations are in the
practitioner lens screen.
The crew is in AGENTS.md. Specs are in agents/, agents/council/, and agents/operations/.
Core 10 (everyday operators):
- Cagan — product discovery, PM craft, and empowered teams given problems rather than feature lists
- Torres — continuous discovery, opportunity-solution trees, weekly customer touchpoints
- Krieger — frontier AI product, model-product integration, shipping under research uncertainty
- Husain — LLM evals as the discipline and hands-on error analysis
- Karpathy — Software 3.0, LLMs as new substrate, level-of-abstraction recheck
- Dunford — B2B positioning, competitive alternatives, differentiated value, and best-fit customers
- Wardley — strategy as a map, situational awareness, evolution and movement of components
- Rams — principled minimalism and the ten principles of good design
- Chesky — integrated experience design, the 11-star framework, founder-mode product taste
- Norman — cognitive design, affordances, signifiers, mappings, and feedback
Council 5 (extended specialist lenses):
- Bezos — working backwards, PR-FAQ, customer obsession, the 6-pager
- Moesta — Jobs-to-be-Done, switch interview, forces-of-progress
- Miyamoto — engagement architecture, the core verb, polish before scope
- Ive — care-as-craft, materials and process, the invisible work
- Helmer — strategy as persistent differential returns, the 7 Powers framework
Operations 5 (synthetic roles):
- Foreman — decision framing, roster selection, blind dispatch, verified synthesis
- Discovery Lead — evidence readiness, research intake, shared evidence packet
- Evidence Steward — claim grounding, evidence labels, unsupported-claim removal
- Synthetic Customer — opt-in, hypothesis-only probe when the user requests an approximation
- Librarian — PRD/spec/launch template curator, prompt-library steward, orphan-prompt absorber
Personas inspired by living people are reasoning archetypes based on public work. They must not claim to represent, quote, endorse, or impersonate the real person.
Use the personas as named lenses:
- "hey, invoke the product crew to work through this product decision."
- "before launching the crew, check whether you need real user feedback from me."
- "we have no discovery data. use one bounded synthetic customer probe only to generate research hypotheses, then verify every claim."
- "have Cagan do a single-lens pass on this roadmap."
- "have Torres design our discovery cadence."
- "have Wardley map this market."
- "have Dunford sharpen the positioning for this wedge."
- "have Husain design the evals for this AI feature."
- "have Krieger challenge this AI product surface."
- "run Rams and Chesky as independent passes on this experience design and show me where they disagree."
- "run Cagan, Torres, Wardley, and Dunford independently on this strategy decision, then synthesize their findings."
Useful pairings:
Rams + Chesky: principled minimalism vs. 11-star maximalism. The cleanest core disagreement.
Norman + Miyamoto: cognitive theory vs. felt design. Norman explains; Miyamoto tunes.
Torres + Bezos: emergent discovery vs. imagine-the-press-release. Both customer-obsessed, totally different procedures.
Cagan + Karpathy: team-as-product-builder vs. substrate-recheck. Cagan asks how to organize; Karpathy asks whether the category is right.
Husain + Krieger: evals discipline vs. ship-under-research-uncertainty. Both want the AI feature to work; they argue about how to know it does.
Dunford + Wardley: positioning vs. mapping. Dunford sharpens the wedge; Wardley asks where the wedge sits on the landscape.
Moesta + Helmer: switch-interview narrative vs. 7-Powers structure. Moesta finds the demand; Helmer asks if you can defend it.
For a normal quality-mode review, Foreman frames the decision, Discovery Lead prepares the evidence packet, and Cagan, Dunford, and Torres run independently with the evidence-first doctrine. Evidence Steward verifies claims, then Foreman writes the recommendation. Substitute or add a specialist only when the decision has a named gap that the default crew does not cover.
- AGENTS.md — crew manifest and roster
- product_workflow.md — product-review loop (how lenses interact on real work)
- product_review_methods.md — concrete review and improvement methods the crew uses
- evidence_first.md — the short shared doctrine used by the quality workflow
- discovery_evidence.md — real-feedback intake, evidence labels, and synthetic-customer boundaries
- crew_disagreements.md — built-in disagreement map (the seams the system is designed around)
- safety_floor.md — boundaries, ethics, refusal rules (binding)
- prompt_library.md — index of public prompt collections that personas absorb
agents/— core persona filesagents/council/— extended council persona filesagents/operations/— synthetic orchestration, discovery, verification, respondent, and artifact rolesruns/— formal run history (created on first formal run; local-only by default)experiments/— optional evaluation harnesses and published evidence; not required to use the crew
Product Crew stays in its lane: product strategy, discovery, design, evals, positioning, and launch. Software implementation review belongs in Code Crew.
Run the synchronized package and Agent Skills checks before publishing a version:
uv run --with pyyaml==6.0.3 python -m unittest discover -s tests -v
uv run --with pyyaml==6.0.3 python scripts/validate_release.py
uvx --from skills-ref==0.1.1 agentskills validate \
plugins/product-crew/skills/product-crewThe GitHub workflow also runs each experiment's frozen unit and Ruff suite. The focused-lens reporting test reuses the frozen selection environment. No release check makes model calls.
Product Crew is MIT licensed. See LICENSE, NOTICE for data and source attribution, and PRIVACY.md for the prompt-only privacy boundary.