ClarifAI
A conversational AI tool that interviews you about a rough product idea and turns the conversation into a structured brief, tracking what’s covered and what’s still open as you talk.
Why I built it
The blank page is the slowest part of any product doc. The fastest way past it isn’t a template, it’s a good interview, the kind a sharp PM or researcher runs. That’s exactly the sort of structured, judgment-heavy conversation an AI can hold, as long as a person stays in the loop to steer and edit.
It also scratches a bigger itch. I design agentic AI workflows for a living, so I want to build them too. ClarifAI let me work through the same questions I face in product design, how to make AI output trustworthy, editable and genuinely useful instead of a black box, with my own hands on the keyboard.
Building it taught me more about designing good AI tools than any amount of speccing would have.
What it does
You start with a rough idea, a sentence is enough. ClarifAI interviews you about it one focused question at a time, plays your answers back, and turns the conversation into a structured brief: problem, target user, solution, success metrics, constraints and scope. You leave with a real draft and a clear list of what’s still unanswered.
- Conversational, not a form. It asks one sharp question at a time and reflects your answer back, so you think the idea through instead of filling fields.
- Coverage you can see. Six topics fill in live as you talk, so you always know what’s covered and what’s left.
- A brief, and its gaps. The output is a structured brief that also names the open questions, instead of papering over them.
- Export where you work. Take it out as a one-pager, PRD, backlog or Figma brief.
⤢ Click to zoomHow it works
⤢ Click to zoomClarifAI runs a real interview, not a questionnaire. You pick the interviewer’s style up front, a skeptical PM, a curious researcher, a structured interviewer or a contradiction detector, and it asks one question at a time in that voice, reflecting each answer back before it moves on. Behind the chat, six topics fill in as they’re covered, so progress is always visible.
⤢ Click to zoomUnder the hood the conversation is progressively summarised, so each turn stays small and cheap, and the model returns structured state rather than just text: what’s covered, the data collected so far, and a flag for when there’s enough to generate. The API key stays server-side, never in the browser. When enough is covered it assembles the brief, and it surfaces the gaps instead of hiding them.
⤢ Click to zoomTech stack
Built and shipped solo, fast with AI assistance and refined by hand. The interview and brief generation run server-side on the Anthropic API, with progressive summarisation keeping each turn cheap.