Your AI Is Building the Wrong Product. Here Is the Prompt That Stops It.

AI doesn’t know what your product is for. It knows what products usually do. That’s why every AI-generated spec looks comprehensive and ships wrong. The Human-Intent Method is the prompt I use to filter out what the model invented and keep what my product actually needs.

When I was building the streak mechanic for AI Advent Challenge, Lovable kept proposing features I never requested. An admin dashboard. A bulk export panel. A notification settings screen. Every suggestion was technically standard for a product in this category. None of them served what I was actually building: a 25-day challenge where the single deciding variable was whether users came back on day two.

I had 48 hours to ship. I cut every proposal that wasn’t streak logic and shipped clean. The features I kept were the ones I explicitly protected. The rest was the model filling in what “standard” looks like for this type of app.

The same thing happens in corporate PM work. You paste a feature brief into Claude, get back a six-section technical spec, and it looks thorough. But half of it is scaffolding the model added because it always adds it: an admin role section, an audit trail module, a notification settings panel.

None of it is wrong. It’s just not yours. The model doesn’t know which of your users needs which of these features this quarter. It fills in what’s plausible and calls it a spec.

Most PMs ship it. The ones who don’t use a prompt like this one first.

This prompt does three things the model never does on its own. First, it strips the generic suggestions from the output. Second, it flags exactly where the AI drifted toward “standard” instead of “strategic.” Third, it rewrites the spec around the specific user friction you defined.

The third step is where the value lives. Most AI critique prompts ask “what’s wrong?” This one asks “what’s wrong for this user’s actual problem?” Generic critique produces generic fixes. Intent-grounded critique produces a spec you can actually ship.

The prompt also forces you to name your user friction before you review the output, which surfaces the decision you were probably skipping over.

Role: AI Product Architect / Filter of Intent.
Context: I am reviewing a feature proposal generated by an LLM.
Task:
1. Review [Insert AI Proposal].
2. Identify 3 "Generic" features that don't serve the core mission.
3. Rewrite the technical spec to focus on [Insert Specific User Friction].
Input: [Insert Feature Draft].

Paste it at the top of a new Claude conversation. Drop your AI-generated spec or PRD into [Insert AI Proposal]. Write one sentence defining your actual user friction in [Insert Specific User Friction]. Something like: “Users abandon the product at step two of onboarding because they don’t see value before they hit a paywall.”

The output is a rewritten spec that cuts what doesn’t belong and protects what does. Run it before you approve anything the model has generated. I run it before I open Lovable for any feature over three screens long.

✅ Use this before any build sprint where AI generated the initial spec.

✅ Use this when a feature proposal “looks right” but feels too broad.

✅ Skip it for UI-only changes where no product decision logic is involved.

Any team can generate a spec in 2026. The person who decides what the spec should actually solve is the only one the model can’t replace. This prompt makes that decision visible and defensible before the build starts, not three days into debugging logic the AI filled in on its own.

If you want help auditing your AI-generated specs for intent drift before they become shipped features, reach out at hello@elenacalvillo.com.