'Story before slides' isn't a new idea. Tome built a whole company on it before shutting down its presentation product in 2025, and STORYD, Chronicle, and Storyflow are all doing versions of it today. So the honest gap isn't 'nobody does narrative-first.' It's narrower than that. Every one of those tools still starts from a blank prompt. You type a topic, they write you a story. None of them take the hundred pages you already have (specs, transcripts, old decks, a pile of Slack threads) and go find the story that's already buried in there.

That's the gap I built this for, originally just for myself. I kept hitting the same wall: the material for a good case study or pitch already existed somewhere, scattered across docs and URLs and screenshots, and turning it into one coherent argument was still entirely on me. The first commit, in May 2026, was a deterministic, no-AI generator, barely more than a template filler. It only became a real pipeline once I started routing that scattered material through actual extraction and synthesis agents instead of summarizing it by hand every time.

A quality gate isn't a feature you add later. It's the actual product. It would've been faster to stop at 'generate a draft' and call it done. Instead there's a dedicated layer before anything renders: a critic that challenges weak claims, a critic that checks pacing, an agent that rewrites flagged beats without touching the rest, an agent that strips the draft of AI-speak so it sounds like the person who wrote it. None of that is optional. A narrative that doesn't sound like you and a slide that clips text at the edge are the same failure: something that looks finished but isn't trustworthy.

Five beats, always. Flexibility was the thing to refuse. Every framework the tool supports (STAR, SOAR, PREP, BLUF, AIDA) still maps to exactly five slots. Users can't add a sixth. That's a deliberate constraint, not a limitation I haven't gotten to. Unlimited flexibility is what makes most AI deck tools produce shapeless output. A story needs a spine, and a spine only works if it can't grow extra vertebrae on demand.

Sixteen agents isn't a feature count. It's an org chart, and someone has to run it. Each subagent does one narrow job on purpose. The actual design work wasn't writing prompts, it was deciding the org chart: where synthesis ends and drafting begins, which stage gets a real reasoning call versus a static rule, where a gate goes before the next stage is allowed to run. Directing that shape, mostly by pairing with AI coding agents rather than hand-writing every layer, is the same skill as running a build. The tool is a case study in the way it was built.

Multi-lens exists because the real fear isn't 'will this look good.' It's 'will I survive the room.' The PM lens, the exec lens, the recruiter lens aren't demo features. They're a rehearsal. Nobody opens a deck tool worried about kerning. They're worried about the one stakeholder who's going to ask the question they don't have an answer for yet. Building the audience simulation before the visual polish was the actual product bet.

The competitive gap wasn't 'nobody does narrative-first.' It was 'nobody starts from what you already have.' Every AI deck tool, including the narrative-first ones, starts from a blank prompt: type a topic, get a story. This one starts from a pile of source material, specs, old decks, screenshots, and finds the story that's already buried in there before a framework is even chosen. That's the actual product thesis, and it's also the hardest agent in the whole pipeline to get right.

Sixteen agents, five layers, one manager deciding what each one is and isn't allowed to do. Source material goes in at one end, a rendered, quality-checked deck comes out the other. Nothing skips the queue.

Intake & Synthesis. Every source document gets extracted in parallel, then narrativeSynthesizer reads all of them together and produces a NarrativeSeed: the recurring themes, the core tension, an evolution arc, candidate angles, and where the evidence across documents actually contradicts itself. intentInterrogator asks 3-5 sharp questions grounded in that seed before any framework gets picked. This layer is the actual product bet: finding the story, not writing one from scratch.

Framework & Draft, then Critique & Audience. creativeStrategist argues for a framework and an angle. intentInterpreter and storyMapper turn that into five beats. Then the murder board runs: a Skeptic critic flags unsupported claims, a Storyteller critic flags dead pacing, narrativeInterrogator targets the weakest beat with a follow-up question, and lensAnalyzer re-frames the same content, unchanged, through a PM's, an exec's, or a recruiter's read.

Quality Gate, then Visual Compilation. Before anything renders, contentFixer surgically rewrites only the flagged beats and voiceCalibrator strips AI-speak so the draft sounds like the person who wrote it. Only then does slideArchitect map beats to layout templates, visualPlanner assign components, and slideInterpreter compile the result into plain semantic HTML that drops into any of 32 themes. A sibling QA app checks the rendered output for accessibility and layout overflow before calling it done.

It needs an OpenRouter API key to do anything AI-powered. Every LLM call routes through one gateway, nothing hardcoded to a single model provider. Without a key, you get the deterministic fallback path this started as.

This is a lab, not a finished product. The 16-agent pipeline (cross-document synthesis, unified critique, voice calibration) is live now, but the newest layer, automated accessibility and layout QA on the rendered output, is the youngest part of the system. Expect the roster and the docs to keep moving.