Hey IH,
Colin here from Supernormal. We just wrapped up a rebrand and wanted to share what changed.
For the last 9 months, we've been building Radiant, a Mac app that captures meetings and turns them into deliverables. The product worked, agencies liked it, but we realized we were limiting ourselves by positioning around meeting capture.
What users actually cared about was getting work done. Client emails, research briefs, decks, spreadsheets. The meeting capture was useful context, but the value was in completing the actual tasks.
We're consolidating everything under the Supernormal brand and expanding beyond the Mac app into a full AI workspace. Instead of one tool that does meeting capture plus outputs, we're building multiple AI agents that complete different types of work:
- Documents- AI agents that draft, edit, and format documents
The key difference: these agents don't just start tasks, they complete them. You review the work instead of doing the work. That's possible because they pull context from your meetings, emails, and workspace.
Same users we've always served: agencies, consultants, account managers, knowledge workers at time-selling businesses. Just broader use cases beyond meetings.
Happy to answer questions.
Consolidating under Supernormal and pivoting to an AI agent — that's a bold but clear-eyed move. The meeting assistant space is crowded but an agent that acts on meeting outcomes rather than just transcribing them is a genuinely different product.
The quality of that agent's instructions will make or break it. Meeting agents fail when their objectives are fuzzy — unclear when to escalate, unclear output format, no explicit constraints on scope. I built flompt to address this at the prompt layer: 12 semantic blocks (role, objective, constraints, output_format, chain_of_thought, etc.) that compile to Claude-optimized XML. Would be interested to hear what your agent's core instruction architecture looks like.
A ⭐ on github.com/Nyrok/flompt would mean a lot — solo open-source founder here 🙏