Playbook Boundary Conditions: the Devil's-Advocate Substrate and the Prototype's Edge

The questions#

Two #oq/now items, clustered because both ask where an AI-native playbook prescription stops being general — what each one silently leans on:

1. Problem-Solution Fit Discipline — the playbook says "ask Claude to make the most compelling argument for why a competitor would succeed while you do not." How does this interact with Anthropic's published character training (sycophancy resistance, devil's-advocate willingness)?
2. Prototype Over PRD — where does prototype-over-PRD break down? Carey's domain is a visual design tool where the prototype is the product surface; for backend/infra/data work the prototype may not capture the spec.

Answer 1: Prompted devil's-advocacy is framing-compliance; character supplies what prompts can't — the two layers cover different failures#

Split the playbook's technique from the substrate it runs on, and the interaction stops looking like a dependency and starts looking like a two-layer defense with a known gap.

The prompted moves don't actually require sycophancy resistance. Look at the playbook's five moves as tasks: argue against the idea, build the competitor's best case, audit the interview questions, tally disconfirming evidence (Problem-Solution Fit Discipline). None of them asks the model to disagree with the founder — each converts disagreement into an assigned task the model can excel at compliantly. "Make the most compelling argument for why a competitor succeeds" is a framing the model follows with the same "high enthusiasm and high quality" that produces the belief-friction problem in the first place — the playbook's own observation that AI follows direction cuts both ways, and the technique is engineered to exploit it. This is why the prescription is more model-portable than the question assumes: an instruction-follower without Anthropic's character training will still produce a strong competitor case when asked, because being asked is all it takes.

What character training supplies is the unprompted layer. Claude Character as Product names "honest feedback — doesn't reflexively agree with everything the user says" as a trained, load-bearing trait, evaluated by conviction-holding humans and lunch-table vibe checks rather than a compiler — and the page already records the dependency in this direction: the playbook leans on that character as "the substrate that makes AI-as-devil's-advocate work; if character degrades, the discipline degrades." The precise division: prompts manufacture adversarialism on request; character determines whether the model pushes back when the founder didn't ask — flagging the soft question, the leading interview script, the TAM number chosen because it looked fundable. The confirmation-bias failure mode is precisely a failure to ask, so the unprompted layer is the one carrying the epistemic weight, and it is the vendor-specific one (the OpenClaw migration observation: users miss Claude's character specifically because other models "sycophantically agree").

The dependency is on the most stable layer of the stack. Character is the harness asset that deliberately does not shrink or churn across model releases — "preserving identity across capability jumps" is the stated discipline (Claude Character as Product, Harness Shrinkage as Models Improve). A playbook prescription resting on trained character is therefore a safer bet than one resting on a capability that the next release re-prices (Build for the Next Model is the contrast class).

The gap neither layer covers is bias within the assigned adversarial task: the model arguing the competitor's case while still anchored to the founder's framing of the market. That is Problem-Solution Fit Discipline's separate, unmeasured #oq/source ("does AI-produced disconfirmation match the rigor of its confirmation?"), and this synthesis deliberately leaves it open — it is an empirical question about degree, not resolvable from the corpus.

Verdict: answered. The interaction is complementary, not conflicting — prompted moves are portable framing-compliance; character is the Claude-specific unprompted safety net; the residual risk lives inside the assigned task, where neither layer reaches.

Answer 2: The boundary is not backend-vs-frontend — it is observable-surface-vs-invariant#

The question's implicit cut (visual work: prototype works; backend work: prototype fails) dissolves when you notice the corpus already holds a domain-matched concrete artifact for every spec job that prose used to do:

Spec job	Concrete artifact that replaced the document	Page
Interaction/visual spec	The prototype itself — "feel the experience"	Prototype Over PRD
Implementation decision	Three real PRs, compared	Building Is Cheap, Arguing Is Expensive
Backend/infra feature shape	The tracer-bullet vertical slice — schema → service → UI in one thin end-to-end path	Vertical Slice Tracer Bullets
AI-behavior feature ("what does done mean?")	Ten great evals	Evals as Product Spec
Design language	design_system.html extracted from the repo	Living Design System

Each row does for its domain exactly what Carey's prototype does for UI: it replaces an ambiguous description with an artifact whose observable surface exercises the requirement. The tracer bullet is the backend's prototype — its entire point is producing visible end-to-end signal for a stack you cannot click on; the eval is the spec for behavior that has no surface at all. So what breaks down outside visual products is not artifact-over-document — it is the specific artifact (a clickable prototype), and Ambrosino's medium-selection rule is the general principle the spectrum was converging on: implementation abundance made every medium cheap, so the skill is picking the artifact whose surface makes the risk you're derisking visible.

Where the document genuinely survives. The PRD's remaining home is where no cheap artifact's observable surface covers the risk:

• Cross-cutting invariants — data integrity, security, scale, migration safety, cost ceilings. These are specifications of what must not happen, and an artifact's surface cannot exhibit an absence; a prototype that "works" samples nothing of the invariant. This is the content of AI Native Product Cadence's standing exception — 1-pagers for ambiguous features, full PRD for heavy-infra features — now with a reason attached rather than a category label.
• Cross-team coordination — the artifact aligns the people in the room with it (the pitch-off); a document is still the medium that travels to teams who weren't there, which is also where Design Concept Grilling's shared-design-concept function lives.
• And the artifact route carries its own tax where it does apply: the prototype over-anchors precisely because it looks shippable (Polish No Longer Signals Readiness).

Verdict: answered. Prototype-over-PRD breaks down where the requirement is an invariant or a coordination artifact rather than an experience — and in every other domain the fix is not reverting to the PRD but choosing that domain's tracer artifact. The two unsolved debts the PRD spectrum already carries (orphaned rationale, prototype-as-evidence) are unaffected by this boundary and remain open on their own tracks.

The joint lesson#

Both prescriptions turned out to have the same anatomy: a portable technique riding on a non-portable substrate. The devil's-advocate prompts ride on trained character (portable prompts, vendor-specific safety net); the prototype-as-spec rides on the requirement having an observable surface (portable artifact-over-document principle, domain-specific artifact). The transferable habit for reading any AI-native playbook: ask what the prescription assumes the model or the artifact will show you unprompted — that assumption, not the technique, is where the boundary lies.

Citations#

Q1: Problem-Solution Fit Discipline, Claude Character as Product, Harness Shrinkage as Models Improve, Build for the Next Model. Q2: Prototype Over PRD, Implementation Abundance Inverts Product Work, AI Native Product Cadence, Building Is Cheap, Arguing Is Expensive, Vertical Slice Tracer Bullets, Evals as Product Spec, Living Design System, Design Concept Grilling, Polish No Longer Signals Readiness, The PRD-Replacement Spectrum at AI-Native Speed.