In this episode, we interviewed Elena Calvillo - a technical product manager who went independent in January 2026 to build AI products full time, and who writes Prompt-Led Product, a newsletter for PMs building in the AI era
Elena spent years shipping software through engineering teams and now ships it herself with prompts - building and running DraftKit, the AI Advent Challenge and CampaignDays, and running technical audits of AI-built products focused on the silent failures demos never show
Her core argument: AI build tools report success without verifying anything actually happened - and verification, the boring audit of the data layer underneath the shiny app, is where the entire risk lives
Key takeaways:
The dangerous failures are silent - a checkout that charges the customer while the database writes nothing is invisible in a demo and fatal in production
Verification is the unfunded risk - leaders fund AI pilots and forget to fund the audit layer, which is exactly where the money is lost
The PRD becomes the operating system - one living spec her AI tools read before every build, so the model never rebuilds from a blank memory
Agent-readability is the next mobile-responsive - within 12 months products legible to AI agents get recommended, and the ones that aren't stop being discovered
🔗 Connect with Elena
Who are you and what do you do?
I'm Elena Calvillo, a technical product manager who spent years shipping software through engineering teams and now ships it myself with prompts. I build and run DraftKit, the AI Advent Challenge, and CampaignDays, and I write Prompt-Led Product, a newsletter for PMs building in the AI era.
What problem did you see that everyone else was missing?
Everyone was debating whether AI writes good code. Meanwhile, the failures I kept hitting were silent ones. My payment checkout worked perfectly while the database wrote nothing. A file upload told 37 contributors their content was saved when it vanished on every single upload. AI build tools will report success without verifying anything actually happened, and almost nobody audits the data layer underneath their shiny new app. That gap is invisible in a demo and fatal in production. PMs are shipping faster than ever with tools like Lovable, which I use daily and love, but speed without verification means you reach the disaster sooner. So I made verification the product: every build I ship gets checked at the database level, and I teach other PMs to do the same.
Senior leaders fund AI pilots and forget to fund verification, and verification is where the entire risk lives.
Walk us through one concrete way your work changes what companies actually ship
Three weeks after I launched DraftKit's collaboration booking system, 43 requests sat in queues, and when I read them, roughly 30% were topic mismatches. Writers were pitching newsletters whose audiences had nothing in common with theirs. Hosts declined manually, requesters got rejected, everyone lost time.
The old workflow: write a spec, pitch it, wait a sprint or three. Instead, I prompt-built the fix in days. An edge function fires while the requester types, pulls both parties' RSS feeds, sends their last 10 article titles to a model, and returns three collaboration topics that fit both audiences before the request is ever submitted. It never blocks a submission, it makes the good pitch the easy pitch.
The person who saw the problem shipped the fix the same week. That is what changes.
What's the most common thing senior leaders get wrong about AI?
They think the risk is AI writing bad code, so they respond with review boards and policy decks. In every audit I run, the expensive failures are boring and invisible: a checkout that charges the customer while the database writes nothing, a daily streak that resets at midnight UTC instead of the user's timezone, a storage function that says "saved" and saves nothing.
Senior leaders fund AI pilots and forget to fund verification, and verification is where the entire risk lives.
The second thing: they still treat builders and deciders as separate people. A PM who can prompt a working prototype in a weekend changes the org chart math, and most leadership teams have not priced that in yet.
My PRD stopped being a document I write for engineers and became the operating system my AI tools run on.
What's in your AI stack? The one tool you rely on every week?
Lovable is my build tool, Supabase is my backend, Claude and Gemini are my auditor and thinking partners. The workflow I rely on every week is treating my PRD as one living document my AI tools read before every build.
Every feature, constraint, and past failure lives in that file, so the model never rebuilds from a blank memory.
My PRD stopped being a document I write for engineers and became the operating system my AI tools run on.
When something ships, the spec updates in the same session. If a behavior is not in the doc, I assume it does not exist. That one habit has saved me more generation credits and rework than any prompt trick I know.
What does your work actually look like day to day?
Mornings are building: one feature in Lovable, then straight into the database logs to check what actually got written, because I trust nothing until I see the rows. Afternoons split between audit work on other people's AI products and writing the newsletter, which forces me to document every build while it is still fresh.
The honest version: around half of my "building" time is verifying things the AI told me were already done.
Stress-testing billing edge cases, chasing a guest invite that never linked to a user account, reading edge function logs. It is unglamorous, and it is exactly why my products stay up. The headline says I ship products with prompts. The real version is I ship, then I interrogate.
Where is your field in 12 months - one specific prediction?
Within 12 months, shipping an agent interface next to your web UI becomes as expected as shipping a mobile-responsive site was in 2015. I already run an MCP endpoint that lets AI assistants query my content library directly, and I publish ai.txt manifests on my products so crawlers like ClaudeBot and PerplexityBot understand what they do.
Products that are readable to AI agents will get recommended, and products that are not will simply stop being discovered.
The money bet: by mid-2027, "do we have an MCP endpoint" shows up in product due diligence the same way "do we have an API" does today. The teams preparing for that now are small, and they will be very hard to catch.
Where should readers find you, and what's the first thing they should join or read?
Start with my newsletter, Prompt-Led Product, at promptledproduct.substack.com. The first thing to read is Part 1 of the Build Series, where I broke my own system in public and documented the year of building that followed.
Everything I claim in this interview has receipts there: the numbers, the failures, and the fixes. I'm also on LinkedIn, where I post build logs between issues.
If you are a PM who knows exactly what to build and keeps being told to wait, that is who I write for.
AI Central Voices is where the AI Central team sits down with the founders, executives, and builders shaping AI - going behind the scenes of how they operate, what they're betting on, and where the industry goes next.
Want to be featured, or have an event you'd like us to cover? Reach out at [email protected]
