Stakeholders in an AI Stack

Stakeholder clarity is not a PM-only skill. It is core Vibe SWE craft, and it matters more now than it did before models could draft. Here is why: when a human writes code by hand, a fuzzy request produces a slow, hesitant build — you go back, ask questions, stall in a good way. When a model drafts, a fuzzy request produces a finished-looking answer in ninety seconds. The ambiguity did not go away. It got faster.

The failure mode

The dangerous failure is not a broken build. It is ambiguity laundered into plausible output. Someone says "we need a settings page," the model produces a full settings page, and it looks so finished that nobody argues with it. Nobody asked what the settings were supposed to do, who asked for them, or whether the ask was ever more than a passing comment in a standup.

The model will not ask "what did they actually mean?" It optimizes for plausibility. It fills your silence with confident output and treats your silence as approval. A senior engineer would push back: "Why does this exist? Who signs off? What happens if we skip it?" The model, as drafted, ships the standup comment straight to main.

What good briefing looks like

Before you hand a request to the model, translate it. Not into a longer prompt — into a clearer one. The fields are simple and none of them are optional:

Outcome: what is true for the user when this ships. Not the feature, the state of the world. Constraints: what cannot change — performance budget, security posture, the deadline. Non-goals: what you are explicitly not building, so the model does not confidently add it. Approvers: who has to say yes — because a model cannot walk down the hall and ask. Risks: where the security and scope landmines are, because the model will not spot the one that matters.

Five fields. Writing them takes ten minutes and replaces an hour of reviewing the wrong thing.

Past craft still counts

Clear communication was never soft skill decoration; it was always the mechanism that turned other people's vague wants into buildable work. The model did not remove that step. It moved it upstream and made it cheaper to skip — which is exactly why skipping it now costs more.

Two other old crafts got more valuable. Judgment about what to push back on: some requests should die in a conversation, not survive into a diff. Ownership of promises: when you tell a stakeholder "it'll be ready Thursday," the model does not sign the promise. You do. The pager still forwards to a human.

The future stack arrives as a faster inbox

You are already living the future stack: several tools running at once, and stakeholder requests arriving faster than you can type. When drafting was the bottleneck, the request queue throttled itself. Now the bottleneck is translation — converting humans into intent at a rate the machine can match.

The engineers who win in this stack are not the ones with the best prompts. They are the ones who ask the stakeholder one more question before the model sees the request. That question is the whole skill.

The loop still runs

Vibe SWE sits under vibe engineering — Willison's term for the senior, accountable practice of tests, evals, and production discipline. This essay is one load-bearing part of that seat: the loop is Build, Ship, Learn, and every stage starts with a human brief the model can actually execute. Build from a stated outcome. Ship with the approver's name on it. Learn what the stakeholder meant, not just what they said.

Read the Ship Loop and what people have shipped on the Ship Wall — then write the brief before the model writes the code. Build. Ship. Learn.

Build. Ship. Learn.