# Spec-First AI PMing, Faster Feedback Loops, and the New Career Signal

*By PM Daily Digest • May 1, 2026*

This issue focuses on a shared pattern across product teams: AI raises the return on strong thinking and exposes weak process faster. It covers spec-first agent workflows, feedback-loop-driven product decisions, communication tactics for PMs, and career signals reshaping the AI PM market.

## Big Ideas

### 1) AI is multiplying process quality, not fixing it
Across strategy, docs, and shipping workflows, the same pattern shows up: AI accelerates whatever operating model is already there. TBM argues that weak practices such as single-player strategy slides and static PRDs become faster and more polished, not better [^1]. Descript's CEO makes the same point from a writing angle: drafting is not just communication, it is thinking, and delegating that thinking to AI weakens downstream decisions [^2]. Gabor Mayer's production workflow reaches the same conclusion from the build side: skipping specification creates context compression, maintainability problems, and dependency failures [^3].

> "Faster bad is still bad." [^1]

**Why it matters:** AI leverage now depends less on prompt cleverness and more on whether your team has sound decision heuristics, living artifacts, and execution discipline.

**How to apply:** Audit where AI is being added as a checkbox. If it is only speeding up stale artifacts or gate-heavy process, redesign the practice first. Use AI to sharpen pre-mortems, prototypes, shared context, and decision summaries instead of automating broken habits [^1].

### 2) Spec-first, multi-agent delivery is becoming a real PM capability
Aakash Gupta argues that the market is rewarding PMs who can show they have shipped AI agents, not just managed AI projects [^4]. The workflow he highlights treats Claude Code as a team: a System Analyst turns requirements into technical specs and tickets, design, ticket, and build work run in parallel, and execution is sequenced through dependency-aware sprints [^3]. In the featured demo, those parallel tracks led to App Store submission in 72 minutes [^3]. The takeaway is not that every PM needs 21 agents on day one; the starting point can be as small as three core roles [^3].

**Why it matters:** This turns AI building from a prototype trick into a skill PMs can use for internal tools, proof-of-work portfolios, and faster iteration.

**How to apply:** Start with a spec-first stack, not a one-prompt stack: System Analyst for requirements, UX Flow Architect for clickable flows, and Spaghetti Agent for code quality [^3].

### 3) Feedback-loop speed is a product advantage in its own right
Granola stayed in stealth for a year so it could change the product before public expectations hardened. That period let the team onboard 150 users by hand, scrap real-time autocomplete because it disrupted meetings, rebuild around calm post-meeting summaries, and cut 50% of features [^5]. AITropos shows a similar pattern at a different layer: the founders spent two years exploring ideas, then moved from waiter hardware to a waiter app to a customer-facing WhatsApp agent before locking onto AI order taking as the wedge [^6].

**Why it matters:** Many product gains come from changing the core interaction or narrowing scope. Those moves are much easier before broad rollout.

**How to apply:** Protect pre-scale learning time. If usage is still small, optimize for faster feedback and bigger changes rather than launch visibility.

### 4) PM communication is becoming multimodal, but the human kernel still matters
Descript's CEO describes a PM workflow built around video: screen-recorded teardowns, short design-review videos paired with Figma or prototypes, launch videos, and AI-generated highlights from long meetings or customer calls [^2]. But she pairs that with a hard boundary: the work should begin with a human kernel of thinking, often captured via dictation, before AI edits for clarity [^2]. She also argues live discussion is still the right mode when the team is in an ambiguous creative stage and needs to "toss the ball around" [^2].

**Why it matters:** Better media does not replace judgment. It changes the bandwidth of how PMs share context and decisions.

**How to apply:** Use AI to compress and polish communication, but create the underlying judgment yourself and keep live conversations for unresolved questions.

## Tactical Playbook

### 1) Replace one-prompt prototyping with a spec-first agent workflow
1. Ask the model what a good system analyst does, then assign that role explicitly [^3].
2. Constrain behavior early: ask clarifying questions one at a time and block documentation until the questions are done [^3].
3. Dictate the full spec, including stack, data rules, security constraints, usage limits, and tone [^3].
4. Generate full Confluence documentation before design or code so every agent works from the same source of truth [^3].
5. Run design, ticket review, and build in parallel: Figma MCP for screens, team review on JIRA tickets, then tagged sprints with manual dependency mapping [^3].
6. Run a code-quality agent after each sprint to catch structural debt before it compounds [^3].

**Why it matters:** This workflow is designed to address the three recurring failure modes of one-prompt building: context compression, unmaintainable code, and dependencies being built in the wrong order [^3].

**How to apply:** Use it when the goal is a production-ready build or a credible PM portfolio item, not just a demo [^3].

### 2) Keep AI in the editing loop, not the thinking loop
1. Dictate the argument you would make live, even if it is rough [^2].
2. Ask AI to tighten the outline and wording while preserving your voice [^2].
3. Do another dictated pass to restore missing nuance or decision criteria [^2].
4. Only publish when the document represents what you actually think [^2].
5. If the team is still exploring, switch from async to a live discussion with a few high-context collaborators [^2].

**Why it matters:** The point of writing is partly to clarify the decision criteria that later make design and execution calls easier [^2].

**How to apply:** Use AI to compress expression, not outsource judgment.

### 3) Evaluate agent systems around one business-critical KPI
1. Pick a single metric that captures whether the agent is doing the job. For AITropos, it is how many order items were identified correctly [^6].
2. Before deployment, run thousands of simulated conversations overnight using customer agents plus analyzer agents [^6][^7].
3. During onboarding, audit live conversations and trigger alerts when something looks wrong [^6].
4. Fix errors manually while patterns are still small, then automate the fix [^6].
5. Keep shrinking onboarding time as domain templates improve [^6][^7].

**Why it matters:** Production reliability in agent systems comes from architecture, evaluation, and feedback loops, not from a good prototype alone [^6].

**How to apply:** Start with the one failure that would break user trust, then build tests and alerts around that first.

### 4) Use AI to remove communication friction before you ask it to replace collaboration
One team in TBM's positive example used AI for status updates, keeping shared context current, and summarizing decisions so people could spend more time in focused 1:1s, better design reviews, and other judgment-heavy work [^1]. Descript's PM workflows point in the same direction: use AI to turn noisy communication into high-signal artifacts, not to avoid the conversation altogether [^2].

**Why it matters:** This is a practical stakeholder-management use case with lower risk than full process replacement.

**How to apply:** Start by automating recaps, summaries, and prep materials. Leave the actual decision-making forum human.

## Case Studies & Lessons

### 1) Granola: use stealth to fix the core interaction before launch
Granola's stealth year was not just about secrecy. It was about increasing feedback-loop speed before public behavior locked in [^5]. The team onboarded 150 users by hand, scrapped a core interaction that pulled users out of meetings, rebuilt around post-meeting summaries, and removed half the feature set [^5]. Hiten Shah called this the key part of Granola's growth story [^8].

**Lesson:** Early growth often comes from subtraction and interaction redesign, not feature expansion.

**How to apply:** If users are learning the wrong behavior, delay scale and fix the workflow first.

### 2) AITropos: prototypes are easy; reliable operations are the real product
AITropos found its wedge only after two years of idea exploration and three product iterations [^6][^7]. The hard part was not building a demo. It was translating messy human conversations into structured POS-compatible data reliably enough for real restaurants [^6]. The team responded with a tools-based architecture for speed, parallelized product searches, pre-fetched context, and fast sub-agents that injected relevant context before the main agent responded [^6][^7].

**Lesson:** In AI products, the durable advantage is often in evaluation and systems design, not the first prototype.

**How to apply:** When a prototype looks impressive, ask what must become deterministic, measured, or parallelized before customers can trust it [^6].

### 3) Descript: treat PM communication as product work
Descript highlights three high-value PM uses for AI video: product teardowns, design reviews, and launch or career videos [^2]. The tool can condense a 14-minute screen-recorded walkthrough into roughly two minutes, smooth edits so they remain watchable, and extract a three-minute highlight reel from a 90-minute meeting or large sets of customer calls [^2]. On the product side, the company tracks how many users export a video on day one; that figure more than doubled over 18 months to roughly one in five users [^2].

**Lesson:** Communication quality is a product surface with measurable adoption, not just an internal hygiene factor.

**How to apply:** If your team already records screens, prototypes, or calls, add an AI editing pass before distribution to raise signal without adding manual work.

## Career Corner

### 1) Shipping an agent is becoming a stronger signal than pedigree
Aakash Gupta argues that 30% of open PM jobs in 2026 are AI PM roles while fewer than 5% of senior PMs have shipped a working AI agent [^4]. He further argues that this gap is letting candidates from non-traditional backgrounds win $1M+ offers at OpenAI, Anthropic, and DeepMind by proving the rare skill directly, though he expects the window to narrow as more PMs ship agents over the next year [^4].

**Why it matters:** In his framing, the market is rewarding demonstrable shipping ability faster than it is rewarding pedigree.

**How to apply:** Build something you can show: an App Store app, password-protected build notes, Confluence docs, JIRA tickets, or agent architecture that makes the work visible [^3].

### 2) The PM-to-CEO path favors founder instinct, but it still has to clear the business bar
Laura's path at Descript ran from IC PM to VP Product to CEO, with the CEO role leaning heavily on founder mentality, product depth, customer understanding, and loyalty to the original vision [^2]. She is equally direct about the trade-off: a product-heavy CEO still has to prove they can drive business outcomes such as stronger margins or customer success, and may need complementary leaders around them while they learn [^2]. Her management strength as VP Product came from hiring exceptional PMs, giving them context, and then enough room to succeed [^2].

**Why it matters:** Product excellence can get you into the CEO seat, but scaling capability determines whether you stay there.

**How to apply:** If you want the path, build both sides: product judgment and the ability to hire, context-set, and operate through others.

### 3) Rewrite the story before it starts showing up in interviews
Deb Liu describes a recurring pattern among recently laid-off high performers: instead of focusing on strategy shifts or market conditions, they narrate the event as a personal failure [^9]. In one example, an exceptional PM came across as guarded and defensive after a difficult previous manager, which cost her an opportunity [^9]. The proposed reset is simple: write the full story, read or listen back to surface the judgment inside it, then rewrite it with less blame and more learning [^9].

**Why it matters:** The story you carry forward affects how you show up and how others read you [^9].

**How to apply:** Do the rewrite before your next interview loop, performance review, or networking cycle.

## Tools & Resources

- [The AI Playbook Puzzle](https://cutlefish.substack.com/p/tbm-420-the-ai-playbook-puzzle): Useful for pressure-testing whether your AI plan is improving the operating model or merely automating it [^1].
- [Gabor Mayer's agent repo](https://github.com/gabormayer/pm): The actual agent files and supporting resources behind the multi-agent PM workflow [^3].
- [Superwhisper](https://superwhisper.com/): Cited in Gabor's workflow as a fast way to dictate dense product specs instead of typing them [^3].
- [Descript CEO episode](https://www.youtube.com/watch?v=YdEkXxZQ0bU): Practical examples of AI-edited teardowns, design reviews, and meeting highlight reels for PM communication [^2].
- [AITropos episode](https://www.youtube.com/watch?v=fFDGEgzcL80): Strong reference for production agent architecture, KPI design, testing, and onboarding in a live operations setting [^7][^6].
- [What is the Story You are Telling Yourself?](https://debliu.substack.com/p/what-is-the-story-you-are-telling): A useful reset for PMs navigating layoffs, difficult managers, or confidence loss before interviews [^9].

---

### Sources

[^1]: [TBM 420: The AI Playbook Puzzle](https://cutlefish.substack.com/p/tbm-420-the-ai-playbook-puzzle)
[^2]: [AI-Powered Video for Team Collaboration | Descript CEO](https://www.youtube.com/watch?v=YdEkXxZQ0bU)
[^3]: [How to Build a Full AI Dev Team in Claude Code | Guide from Google PM Gabor Meyer](https://www.news.aakashg.com/p/claude-code-dev-team)
[^4]: [substack](https://substack.com/@aakashgupta/note/c-251919242)
[^5]: [𝕏 post by @IvanLandabaso](https://x.com/IvanLandabaso/status/2049785983814472096)
[^6]: [Building AI Employees for Hospitality: How AITropos Takes Orders Where Customers Already Are](https://www.youtube.com/watch?v=fFDGEgzcL80)
[^7]: [𝕏 post by @ttorres](https://x.com/ttorres/status/2049900091134111786)
[^8]: [𝕏 post by @hnshah](https://x.com/hnshah/status/2049900185195839884)
[^9]: [What is the Story You are Telling Yourself?](https://debliu.substack.com/p/what-is-the-story-you-are-telling)