the wedge / new for 2026

Feasibility is free. Viable and lovable is the bar.

Most AI PM prep bolts "AI" onto old frameworks. This does not. When anyone can build anything, interviews test whether you can find a problem worth paying for and a product people actually love.

Building is no longer the rare signal. Judgment is. AI made it possible for a PM to prototype a working product in an afternoon, so the interview moved upstream and downstream. Upstream, is this a problem worth solving, for people or companies willing to pay? Downstream, can this become something people trust, return to, and prefer over the way they work today?

That is the bar this section takes seriously. Viable means paid pain, honest market size, margin, and a right to win that survives a capable clone. Lovable means workflow fit, trust, restraint, recovery, and the discipline to leave the model out when it would make the product worse. A clean demo can impress for a minute. A product that earns repeated use has to survive cost, latency, mistakes, edge cases, and human impatience.

The AI PM interview now tests those muscles directly. Eval harness questions ask whether you can define quality before users find the failures. Vibe-coding rounds test what you choose, cut, narrate, and defend under pressure. Guardrail prompts test whether you understand irreversible actions, spend limits, human confirmation, and kill switches. Unit economics questions test whether your product can afford its own intelligence. "When the AI is wrong" tests whether you can protect trust after the demo breaks.

Use these guides as a map through that new loop. The goal is not to sound fluent in AI vocabulary. The goal is to answer like someone who can own the product quality decision: when to build, when to stop, what to measure, where the risk lives, and why a user or buyer would still care after the novelty fades.

Agent guardrails cheat sheet for PM interviewsA four-point rubric for answering "design guardrails for an agentic AI" questions, with concrete examples and interviewer scoring criteria. read → Anthropic PM interview values: what to read and what the round actually testsThe required reading list and scoring rubric for Anthropic's 45-minute culture round, with strong/weak answer examples. read → Build an eval portfolio project (the 48-hour version)A step-by-step guide to building an AI eval project you can point to in interviews, from scoping to golden dataset to published results. read → How to answer a cost-per-query question in an AI PM interviewA worked arithmetic answer for the AI PM unit-economics question. Real 2026 token prices, the reasoning-token trap, and the viability gate. read → The day-two playbook: what a PM does when the model regressesA concrete triage sequence for AI model regression in production: four regression types, rollback decisions, and the interview answer that clears the bar. read → Eval harness PM interview answer: the structure that clears the barThe exact answer structure for "walk me through your eval harness" in an AI PM interview, with strong/weak examples and the components that signal seniority. read → Build an eval harness in an afternoon (no engineering)A PM-executable guide to building a working AI eval harness: golden sets, LLM-as-judge calibration, ship gates, and the flywheel logic behind it. read → Feasibility is free: why viable and lovable are the barThe core 2026 AI PM thesis. When anyone can build anything, the PM job is proving viability and making something people love enough to return. read → Frontier lab comp decoded: PPUs, RSUs, and the liquidity thesisHow OpenAI PPUs, Anthropic double-trigger RSUs, and tender offers work for PMs evaluating or negotiating a frontier lab offer. read → How AI changed what PM interviews test in 2026Feasibility collapsed, usability has a floor. PM interview loops grew to 4-6 rounds testing viability, lovability, eval literacy, and live prototyping. read → How interviewers catch AI answers in a PM loopFour PM-specific tells interviewers use in 2026 to spot AI-washed candidates, from metric quality to the degradation drill. read → How our grader scores PM interview answersThe full rubric, bias mitigations, and 2026-specific criteria behind our AI grader. Five dimensions, behavioral anchors, no black box. read → When not to use AI: killing an idea as the senior moveA kill-decision framework for AI PMs. When to stop an AI feature, what signals to name in the interview, and how to frame restraint as judgment. read → LLM unit economics: the one-pager every PM needsToken cost formula, 2026 model prices, prompt caching, and model routing explained so you can reason through the math live in any interview. read → Lovable, not just usable: the 2026 product barUsability has a floor now. What lovable means in practice, how to measure it, and how interviewers probe for the judgment of when not to act. read → The obnoxious AI antipatterns catalogueA named taxonomy of AI features that erode trust, trap users, and destroy lovability. For PMs building, removing, or interviewing about AI products. read → Proving viability when anyone can build itHow to answer "should we build this?" in 2026: paid pain, honest market sizing, and a right-to-win that survives the commodity model question. read → RAG vs fine-tuning vs prompting: how to chooseThe PM decision ladder for AI architecture. When to use prompting, RAG, or fine-tuning, with cost benchmarks and the failure modes interviewers probe. read → Should a model even be here?The pre-design gate every AI PM must own in 2026. Four decision tests for whether a problem warrants a model at all, with strong and weak interview answers. read → AI tools in PM take-home assignments: the company policy matrixWhich companies ban AI in PM take-homes, which require it, and how to use it without sinking your submission or the debrief. read → How to pass the vibe-coding round when you can't codeThe PM vibe-coding round tests judgment, not syntax. A 45-minute framework, tool selection guide, company-specific scoring, and the debrief moves that close strong. read → Which vibe coding tool to pick for your PM interviewA decision guide for PM candidates choosing between v0, Lovable, Bolt, Replit, and Cursor for a 45-minute prototyping round. read → When the AI is wrong: the PM interview answerHow to answer the "your AI gave a confident wrong answer" interview question. In-the-room recovery, blast radius scoping, and trust architecture design. read →