Once AI writes the code, the real moat is deciding what to build — and the toolkit is a set of old trades worth picking up on Monday.
You Can't Prompt the Room: The Last Skill AI Won't Replace - Balázs Horváth, VisualLabs · Balázs Horváth
16 min total·Actually worth watching closely: ~3 min·1 must-watch clip
- 0:00 – 2:10Listen
Opening: 21 agent ideas, 4 survivors
The speaker opens with real numbers from an internal hackathon: of about 21 agent ideas, 17 were abandoned because there was no data access or no business value, and only 4 had a big impact on how the team works today.
When building capacity is abundant, the waste isn't in writing the code — it's in picking the wrong thing to build.
A purely spoken opening story with no visual dependency; the numbers and the conclusion are all you need.▶ Jump to 0:00Speaker · Balázs Horváth - 2:10 – 4:20Listen
The bottleneck moves: you can't prompt the room
The argument that once code is no longer the bottleneck, the bottleneck becomes getting stakeholders and decision makers into the room and eliciting the requirement — and when everyone has the same models, the difference is who reads the room.
You can prompt your code, your AI, your whole specification — you can't prompt the room.
The core argument; no visual dependency, pure exposition, fine to listen to like a podcast.▶ Jump to 2:10Speaker · Balázs Horváth - 4:20 – 7:04Watch
The analyst toolkit and the user story map
Proposes that the new core skill is the functional consultant's old trade — story mapping, business model canvas, value canvas — and demonstrates on a support system how a story map carves out the MVP and the backlog.
The first row of user stories is your MVP; the second row goes to the backlog — do a mapping session before you build.
The story map visual at 279s is worth pausing on: the backbone, the lanes and the two rows of stories are spatial, and the layering doesn't come through on audio alone.▶ Jump to 4:20Speaker · Balázs Horváth - 7:04 – 10:20Listen
Value first, and the four questions before you build
Introduces the order of thinking — start from value, understand how value is created, then design the system — and gives the four questions to answer before building, with the answers written into a markdown file in the repo that AI can read.
Four questions: whose problem is this, what does winning look like, what would make them refuse to use it, and would it change a decision — with the answers kept in the repository for AI to access.
The framework and question list unfold verbally with no flagged visual moment; writing the four questions down beats watching the slides.▶ Jump to 7:04Speaker · Balázs Horváth - 10:20 – 14:08Listen
Feeding AI standard stories, and the anti-pattern alarms
Explains why packaging standard-format user stories together with acceptance criteria gets better results from AI, and lists what building the wrong thing looks like, plus the faster-horse trap.
AI is coded to give the most common answers, so using it only to improve what exists replicates a faster horse; go back to the structure AI knows and it gets you better results.
The template and the anti-pattern list are spoken, with nothing demonstrated on screen; use the alarms — velocity up but adoption poor — as a self-check while listening.▶ Jump to 10:20Speaker · Balázs Horváth - 14:08 – 15:43Listen
Closing: two things to start Monday
Ends with immediately actionable moves: replace "features shipped last quarter" with features actually used more than twice, and move the real subject matter experts where they have impact on what gets built.
Audit what you track, measure the features that get reused, and involve the most experienced people in deciding what should be built — so you build the right thing and not just the next thing.
No visual dependency; it's a call to action, so listen through and note the two to-dos.▶ Jump to 14:08Speaker · Balázs Horváth