Home
May 13, 2026
View All

Teresa Torres Builds AI-Generated OSTs; GitHub Copilot Reprices

·1 underrepresented voice

The Short Version#

Teresa Torres used vibe coding to build AI-generated Opportunity Solution Trees for her own product — and the behind-the-scenes writeup is one of the most honest PM-as-builder dispatches we've seen. Meanwhile, GitHub Copilot restructured its individual plans with flex allotments and a new Max tier, and Vercel shipped natural-language firewall rules — a small feature that signals something larger about where infrastructure UX is heading.

Teresa Torres — Behind the Scenes: Building AI-Generated Opportunity Solution Trees#

Source: https://www.producttalk.org/behind-the-scenes-ai-osts/ Credibility: High (first-party writeup from Teresa Torres, published on her own site)

What happened: Teresa Torres — creator of the Opportunity Solution Tree framework and one of the most influential voices in product discovery — just finished what she calls "an all-out engineering sprint" to build AI-generated OSTs into her own product. She's explicit that she doesn't think of herself as an engineer, which makes this dispatch particularly interesting: it's a PM-turned-vibe-coder building a core AI feature for a tool that teaches PMs how to do discovery.

Key patterns:

  • Torres is using AI coding tools to ship features she couldn't otherwise build herself — this is the PM-as-builder pattern playing out at the craft level, not just as a concept
  • The feature she's building (AI-generated Opportunity Solution Trees) is a direct application of her own framework — she's dogfooding the hardest possible use case
  • The post is behind a paywall for the audio version but publicly accessible, suggesting she's treating this as both product development and audience engagement

Why it matters for PMs: Two things are happening here simultaneously. First, the product itself: AI-generated OSTs represent a significant test of whether LLMs can meaningfully assist with product discovery frameworks — not just surface-level templates, but the kind of structured thinking Torres has spent years developing. If it works, it validates a whole category of "AI coach" products. Second, the process: Torres is a non-engineer who just did an engineering sprint. That's the vibe coding thesis made concrete by someone whose judgment PMs trust. How she describes what worked and what didn't is worth reading carefully.

Critical questions:

  • How does the AI-generated OST handle the most common failure modes Torres teaches against — solution-first framing, vague opportunity statements, weak connections between levels?
  • What model(s) is she using, and did she have to do significant prompt engineering to produce OSTs that meet her own quality bar?
  • Is the output a starting point users edit, or a finished artifact? The trust and control question is directly relevant here.
  • How does she plan to validate that AI-generated OSTs actually improve discovery outcomes for users, rather than just reducing friction?

Action you could take today: If you use OSTs or continuous discovery in your practice, read this post and map Torres's AI feature decisions against your own team's most painful discovery bottlenecks — the places where you wish you had a starting point.

GitHub Copilot — New Individual Plan Structure: Flex Allotments, Pro+, and Max#

Source: https://github.blog/news-insights/company-news/github-copilot-individual-plans-introducing-flex-allotments-in-pro-and-pro-and-a-new-max-plan/ Credibility: High (first-party announcement from GitHub blog)

What happened: GitHub restructured its Copilot individual plans. The new structure introduces "flex allotments" in the Pro and Pro+ tiers — a usage model where you get a base allotment that flexes based on how much you use — and adds a new Max plan at the top of the stack. This is a meaningful pricing architecture change, not just a tier rename.

Key capabilities:

  • Flex allotments in Pro and Pro+: Users get a base allocation that scales with usage rather than hard caps — this is a shift away from "use it or lose it" monthly limits
  • New Max plan: Positioned above Pro+, targeting power users who need higher throughput or more model access
  • The restructuring applies to individual (not enterprise) plans, suggesting GitHub is trying to capture more revenue from high-usage individual developers before they churn to competitors

Why it matters for PMs: This is directly relevant to anyone thinking about AI tool pricing strategy. GitHub is making two bets here: (1) that flex allotments reduce the "I hit my limit" frustration that drives churn or competitor comparison, and (2) that there's a segment of individual users willing to pay significantly more for a Max tier. The second bet is interesting — it implies GitHub has usage data showing a long tail of power users who are currently underpriced. If you're pricing any AI feature, this is the pattern to watch: start with hard caps, identify power users through usage data, then create a higher tier specifically for them.

Critical questions:

  • What exactly is a "flex allotment" — is this metered overage (pay per use above base), or a soft cap that expands automatically?
  • How does this change the effective price for heavy users compared to the old Pro plan?
  • Is the Max plan positioned to compete with Cursor's pro tier, or is it targeting a different use case entirely?
  • Does this affect the Business/Enterprise plans, or only individual?

Action you could take today: Pull your team's Copilot usage data and identify who's hitting allotment limits regularly — those are your Max plan candidates, and they're also the users most likely to evaluate Cursor or Windsurf as alternatives.

Vercel — Create Firewall Rules with Natural Language#

Source: https://vercel.com/changelog/create-vercel-waf-custom-rules-using-natural-language Credibility: High (first-party changelog entry)

What happened: Vercel shipped natural-language creation for Web Application Firewall (WAF) custom rules. Instead of writing firewall rule syntax, you describe what you want to block or allow in plain English and Vercel generates the rule. This landed alongside two other firewall-related updates this week: CLI management for firewall rules and Trusted Sources for Deployment Protection.

Key capabilities:

  • Natural language input → WAF rule generation (covers custom rule creation, not just presets)
  • Complements the new CLI firewall management shipped the same week
  • Part of a broader pattern: Vercel is applying natural language interfaces to infrastructure configuration, not just application code

Why it matters for PMs: This is the "AI in the plumbing" pattern that's easy to overlook but matters for product strategy. Vercel is making security configuration accessible to developers who don't know WAF syntax — which expands who can actually manage their own infrastructure. For PMs: this lowers the barrier to shipping secure products without a dedicated security team, which is meaningful for small product teams. More broadly, it's evidence that natural language interfaces are moving beyond chat and code generation into configuration management. The question of "where does NL input make sense vs. where does it introduce risk?" is worth thinking through for your own product.

Critical questions:

  • What's the accuracy rate on generated rules? A miswritten firewall rule could block legitimate traffic — how does Vercel validate the output before applying it?
  • Is there a review step before rules go live, or does the AI-generated rule apply immediately?
  • How does this interact with the existing preset rules — can NL rules conflict with base protections?

Action you could take today: If your team manages Vercel deployments, try the NL firewall rule creator on a real protection scenario you've been putting off because the syntax felt opaque.

Quick Hits#

The Thread#

The PM-as-builder pattern is becoming concrete. This week: Teresa Torres doing an engineering sprint to ship AI-generated OSTs, Vercel making firewall configuration accessible to non-security engineers via natural language, and GitHub restructuring pricing to capture the long tail of power developers. The throughline is that AI is collapsing the distance between "person with the context" and "person who can ship the thing." The interesting PM question is no longer "will this happen?" — it's "what breaks when the person with context and the person who ships are the same person?"

Sit With This#

Teresa Torres — the person who literally wrote the framework for structured product discovery — just spent a sprint trying to get AI to generate Opportunity Solution Trees that meet her own quality bar.

For your product: Pick one core framework or structured output your team produces regularly (a PRD, a user story map, a jobs-to-be-done interview synthesis). What would it take for an AI-generated version to be "good enough to edit" rather than "wrong enough to redo"? What's the quality bar, and who in your organization actually knows how to evaluate it?