# Handoffs Collapse, Evals Rise, and PM Work Gets More Technical

*By PM Daily Digest • May 29, 2026*

This brief covers the collapse of traditional PM-design-engineering handoffs, practical patterns for building eval-driven AI products, and the career shift toward more technical, founder-like product work.

## Big Ideas

- **AI is collapsing the PM-design-engineering relay race.** At OpenAI, PMs write markdown PRDs that Codex turns into shipped PRs; designers ship full UIs with instrumented noop backends; engineers focus on constraints, review agents, and build-failing tests in a shared repo [^1]. Gokul Rajaram’s “pod-of-one builder” describes the same direction: AI compresses execution, so the scarce skill becomes judgment—choosing the right problem and spotting mediocre output [^2]. **Why it matters:** coordination work is shrinking; product leverage is moving toward taste, validation design, and system constraints.
- **The interface shift is from task-based work to agent orchestration.** Emma from Resonant says PMs are becoming “agent orchestrators” who train agents over time so autonomous workflows do not turn into “AI slop” [^3]. Scott Belsky makes the product-level version of the same point: replace task-based workflows with ask-based workflows at the OS level, while keeping underlying models swappable beneath the UI [^4]. **Apply it:** encode principles, examples, and guardrails in reusable artifacts agents can act on.

## Tactical Playbook

- **Use a painted door before you build the backend.**
  1. Ship a real UI with a noop backend [^1].
  2. Instrument clicks and flows to see where demand is real [^1].
  3. Build APIs only where behavior justifies the investment.

  This gives product and design real evidence instead of speculative prioritization.
- **Start AI products with evals, not giant prompt boxes.** Lorikeet is moving from “give us your SOP” toward first defining what good and bad outcomes look like, then having a Coach agent generate SOPs, guardrails, and test cases [^5]. When the model is close but missing one fact, they use “resolution in the loop”: the AI pauses for targeted human input instead of escalating the whole ticket [^5]. **Apply it:** define failure cases, escalation rules, and knowledge gaps before you tune prompts.

## Case Studies & Lessons

- **Lorikeet followed customer pull, not founder intuition.** The team spent months on reflection tools and information dashboards before a healthcare startup made the real job explicit: help clear the support inbox [^5][^6]. Their first prototype was a command-line workflow using real customer CSVs, which let them iterate quickly on real data [^5]. Today the product runs a ticket-handling Concierge agent plus a Coach agent for configuration and improvement [^5][^6]. At scaled customers, human average handle time went **up** because people were now spending more time on the hardest tickets [^5]. **Takeaway:** good AI automation can increase the value of human work rather than simply reduce it.
- **AI can compress enterprise discovery into traceable artifacts.** In the ACNA example, a messy **$2.8B** settlement with complex eligibility rules became a “First Pass” pre-validation layer that checks completeness before claims reach reviewers [^7]. The workflow then turned those rules into decomposed user stories, test cases, and a living playbook [^7]. **Takeaway:** in dense legal or operational domains, AI is most useful when it makes ambiguity auditable.

## Career Corner

- **Founder skills are becoming PM skills.** Emma argues PMs are increasingly founder-like and well positioned to start companies [^3]. The gaps she highlights are practical: getting from prototype to credible demo, understanding fundraising dynamics, and choosing between sales-led and product-led growth [^3]. Her view: mildly technical PMs can now get to a real customer demo with tools like Lovable, and sales-led motions may be an easier way to get early traction than pure growth hacking [^3]. **Apply it:** build one demo yourself and be ready to talk about distribution, not just features.
- **AI PM interviews are getting more technical.** Exponent reports that Meta’s AI product sense round combines a traditional 30-minute product case with 30 minutes of live prototyping in Llama, followed by questions on token efficiency, latency, retrieval, and compute trade-offs [^8].

> "The one thing I told them to remember above all in the interview was: tell them what you’ve learned about users." [^9]

**Apply it:** prepare both a product narrative and a technical one: what you would build, what data it needs, where performance trade-offs show up, and use real data sources when possible [^8].

## Tools & Resources

- **Product Playbook pattern for AI discovery.** The workflow moves from freeform idea exploration to vision, personas, JTBD, event flows, user stories, test cases, and lean canvas outputs, then stores them in a living “Product Playbook” that teams can revisit as markets change [^7]. Worth exploring if your team keeps losing context across meetings and stale docs.
- **Lovable for fast customer demos.** Emma cites it as a practical bridge from rough prototype toward something a mildly technical PM can show customers for feedback, even if it is not yet production-grade [^3].

---

### Sources

[^1]: [substack](https://substack.com/@aakashgupta/note/c-266868869)
[^2]: [𝕏 post by @sachinrekhi](https://x.com/sachinrekhi/status/2060013259101716562)
[^3]: [Lessons from Shifting PM to Founder | ProductTank London](https://www.youtube.com/watch?v=swzrQiw1dYs)
[^4]: [𝕏 post by @scottbelsky](https://x.com/scottbelsky/status/2059997457820295193)
[^5]: [Building Lorikeet: How AI Humility and a Dual-Agent Architecture Are Redefining Customer Support](https://www.youtube.com/watch?v=eZj1xSiyd9U)
[^6]: [𝕏 post by @ttorres](https://x.com/ttorres/status/2060046952549536048)
[^7]: [Building Products That Achieve Org Outcomes with AI & Human Experts | Rezoomex](https://www.youtube.com/watch?v=MhxhBmulrHo)
[^8]: [What REALLY Happens in Meta’s AI Product Sense Round](https://www.youtube.com/watch?v=ipFrkkw-CgY)
[^9]: [𝕏 post by @paulg](https://x.com/paulg/status/2060078726109442266)