A collaborative AI workspace for marketing, product, and sales teams. Generate personas in minutes, interview them like real customers, and export them into a codebase when the work moves downstream.
Quikest.ai is a client SaaS we designed, built, and continue to operate. A multi-tenant workspace that turns a five-field product definition into a set of personas a team can hold in common — edit together, interview, and travel with into the codebase. Live since 2024.
Client SaaS · commissioned, designed, and shipped by Creative Mantra. Multi-tenant workspace platform.
Real-time collaborative editing, AI-generated portraits and environments, agent-context export.
Product, design, and engineering — delivered end-to-end by Creative Mantra.
7-day trial · Free and Pro plans. Multi-workspace, permissioned, Stripe-billed.
“We built Quikest because we kept watching the same failure mode. A marketing lead, a PM, and a sales engineer would each say the customer in the same meeting and each mean someone different. The disagreement about the headline was, underneath, a disagreement about who was reading it. Quikest is what we built to end that argument.”
Three quiet failures every cross-functional team already knows. Set out here plainly, so the rest of the page reads against them.
We built Quikest to make the shared reference a thing — one document, edited together, alive.
Every cross-functional team has this quiet problem. Nobody is wrong. Nobody is aligned. The disagreement about the roadmap is, underneath, a disagreement about who the roadmap is for.
A PDF, a Miro board, a static slide. It gets filled in once, shared in a launch deck, and then goes stale by the second sprint. Nobody edits it. Nobody talks to it.
The reference exists in fragments. It is not in one place. It is not editable together. And it is not something a fourth person joining the team next week can hold in their head.
“The interviews reshaped three product decisions that we then held to. Everything else on the page follows from these.”
The moment a persona sits in one person’s Notion and nobody else can touch it, it starts dying. Real-time collaboration is the fabric, not a feature.
Reading a persona bio is not the same as asking the persona a question. The Ask AI sidebar exists so teams can pressure-test a claim against a live reference.
For engineers that means: into the codebase, as context the AI assistant sees. That is what the .claude/ export is.
A persona is not a form. It is a shared reference — the one the marketing lead will point to when she says she wouldn’t click that headline, the one the product manager will hold in mind when he’s picking between two features, the one the sales engineer will name in a demo to a room of skeptics. Quikest is where that shared reference gets made.
The system separates four surfaces by intent — the intake surface (five fields or a URL scan), the synthesis pipeline (two LLM stages plus two image models), the collaborative editor (TipTap on Y.js), and the export surface (.claude/, .cursor/, generic) — and unifies them through a single persona document held in Supabase. Real-time is Y.js + Hocuspocus. Auth is Clerk. Files ride Vercel Blob. LLMs are routed through the Vercel AI SDK.
A two-stage LLM pipeline plus two image models — small and fast for the wide-net stage, measured for depth, portrait and environment generated in parallel and cached in Vercel Blob. All unified through the Vercel AI SDK.
A team starts one of two ways. They paste a product URL and Quikest scans it — extracting product name, industry, target audience, description — or they fill in five fields by hand. A lifecycle stage picker (Concept through Decline) sits on top.
The intake goes to Gemini 3 Flash Preview via OpenRouter. The output — three to five persona archetypes, each a name, a category, and a short description — streams back to the frontend in real time. Nothing is generated the user did not ask for.
Each promoted archetype goes to GPT-4o-mini for demographics, psychographics, motivations, objections, and a TipTap-ready biography. Gemini 2.5 Flash Image generates the 1:1 portrait. Gemini Imagen 4.0 generates the 16:9 environment. Both cache in Vercel Blob. Both can be regenerated.
“We picked the small model for the wide-net stage, and the measured model for the depth stage. That is not a hack; it is the shape of the problem.”
— internal design review, 2024
Quikest routes every LLM and image call through the Vercel AI SDK — a provider-agnostic abstraction that turns “switch from OpenAI to Google” into a one-line change instead of a rewrite. The routing is by job shape, not by vendor preference.
| Job | Provider · Model | SDK | Output shape |
|---|---|---|---|
| Persona recs | Google · Gemini 3 Flash Preview | @openrouter/ai-sdk-provider | 3–5 archetype JSON, streamed |
| Full profiles | OpenAI · GPT-4o-mini | @ai-sdk/openai | structured JSON + rich-text bio |
| Portrait (1:1) | Google · Gemini 2.5 Flash Image | @ai-sdk/google-vertex | PNG · cached in Vercel Blob |
| Environment (16:9) | Google · Imagen 4.0 | @ai-sdk/google-vertex | PNG · cached in Vercel Blob |
Every AI call is metered in tokens (Tiktoken) and billed against a workspace credit balance. When a workspace runs out of credits, the isAiAllowed check quietly blocks the request. No mystery bills.
A persona described only in text is a résumé. A persona with a face and a setting is someone the team can picture across a desk. Quikest generates both — and both regenerate on demand. Portraits come from Gemini 2.5 Flash Image; environments from Gemini Imagen 4.0. Both cache in Vercel Blob so the CDN handles retrieval, not the LLM budget.
Portraits and environments are stored together as a single persona artefact — the person and the room they sit in. That’s how the product renders them, and how this page presents them: not two files, one reference.
Ask AI sits in a sidebar on desktop, a modal on mobile. Would you pay for this? What would make you cancel in the first week? What headline would stop your scroll? No prompt engineering. Send.
The response comes back in-character. Where the answer makes a claim that has a source, source references — links, images — attach beneath the response.
Not generic follow-ups — the questions the persona would most likely raise next. The team is not drafting the next question from cold. This is the reference talking back.
“The persona started to feel less like a document and more like a colleague we could tap on the shoulder.”
— early-access user, 2024
Every persona in Quikest is a TipTap v2 document — ProseMirror underneath, Y.js binding on top, TipTap Pro AI extensions for slash-command completions and structured export. Which is a technical sentence for what it feels like to use: a Notion-caliber editor with real-time sync, real cursor presence, and an AI menu that respects the persona context you’re editing against.
/rewrite, /expand, /simplify · tone from 14 options · language from 13.
Prompt-aware completions bound to the persona document context.
Notion-parity editing primitives inside the persona surface.
Tables, TOC, DragHandle, Typography — the base extensions.
Structured biography sections. Jobs-to-be-done as checkable tasks.
Semantic block nodes that flatten cleanly on DOCX export.
One-click download. Clipboard-copy Markdown as a second path.
Renders the persona document to a Word file the wider org can open.
The polish extensions that make a persona document read like a designed page.
Persistent formatting the CRDT tracks alongside the text.
The writing-craft extensions. Small features. Big effect on drafts.
Enable per-workspace. Off by default. Writers ask, we ship.
Underneath the editor is a Y.js CRDT — a data structure that resolves concurrent edits at character-level granularity without asking either author which one wins. TipTap v2 syncs through a Y.js binding to a Hocuspocus WebSocket server. Every open document is a room. Every teammate in the room gets a color-coded cursor, a name tag, and a real-time view of the other cursors moving. Comment threads ride Supabase Realtime alongside the CRDT.
Chris wakes at 5:14 am, checks HRV on his wristband, and picks the day’s session from the plan he built with his coach on Sunday. Breakfast is a bowl of oats and a black coffee, always in the same ceramic mug— Maya
He treats gear as instruments, not fashion. Two pairs of the same shoe, retired at 480 km. His objection to any subscription is speed — if the app takes more than one screen to log a session, he uninstalls by the second week.
Motivations: control over the plan, clean data, and a coach who reads the numbers rather than the vibes. He will pay for tools that respect his time; he will churn from tools that ask him to fill in what they should have inferred. Devi
The Y.js document is mirrored to IndexedDB in the browser through y-indexeddb. When the WebSocket to Hocuspocus goes away — an airport, a tunnel, a hotel — the editor keeps writing to local storage. When the connection comes back, the CRDT rebroadcasts the offline delta and the other clients converge. There is no “reconnect” dialog. The header sync chip goes from offline back to synced and the work continues.
A persona in Quikest is a shared reference. Shared responsibly. The permission model is three tiers — Owner, Write, Read — and the tiers exist at the persona and document level, not just the workspace level. Workspace members can be invited from the directory; external users can be invited by email, but only by an admin. An org-wide toggle sets a default. A shareable link copies the persona URL for the everyday case where you just want someone to see the thing.
Row-level security in Supabase enforces every read. Clerk provides the JWT that Supabase verifies. There is no permission logic in the frontend that the backend does not re-check.
Auto-discoverable by each target · Client-side ZIP via JSZip · No workspace data leaves the browser during export
Which is fine if the persona lives on a wall in a marketing pod. But the person building the feature the marketing pod was pointing at is an engineer, and their most-used surface is not a slide — it is a code editor with an AI assistant sitting inside it. If the persona does not travel there, the persona does not travel.
Three targets: .claude/ for Claude Code, .cursor/ for Cursor, generic/ for GitHub Copilot and anything else that reads a nearby text file. The ZIP is generated client-side (JSZip) so nothing leaves the browser during export.
The engineer never has to configure a thing. The persona rides along — the same reference the marketing lead was pointing at, now sitting next to the code, understood by the assistant that helps write it.
“We built the export in a week. It has been the single most-quoted feature by engineering teams since.”
— internal review, 2024
Every generation — persona recommendations, full profiles, portraits, environments, Ask AI turns, editor slash-commands — is metered in tokens using Tiktoken and billed against a per-workspace credit balance. The Subscription page (admin-only) shows the balance, the burn rate, and the plan tier. When the balance runs out, the isAiAllowed check quietly blocks further AI calls and a soft upgrade prompt appears in context. No overage bills, no mystery charges.
Limited personas, limited documents, limited AI tokens. Enough to run one product definition through the pipeline end-to-end.
Higher limits across personas, documents, and AI tokens. Multi-seat. Admin permissions. Stripe-managed subscription.
Every new workspace gets a week of unrestricted usage on signup. No card required to start.
Payments are Stripe. Subscription management is the Stripe portal. Invoices download from there. Nothing bespoke.
Quikest’s design system defines fifteen steps of sage green, from a near-white #F8FAFB down to a nearly-black #2D342A. Production uses one. --sage-12 in light mode. --sage-8 in dark mode. That restraint is not a design accident — it is the product’s opinion about what a workspace should feel like: quiet, considered, out of the way of the work. A persona document is the loudest thing on the screen. Everything else recedes.
Instrument Serif for occasional weight. Inter and Manrope for body. Radix + shadcn/ui underneath. No mood-board neon. No “dark mode with a purple accent.” Sage green, gray text, black type. That’s the palette.
Every dependency below is in production. Every version is verified in package.json. Nothing here is aspirational.
Not in the stack: no vector database, no MCP servers, no multi-agent framework, no LangChain, no LangGraph, no fine-tuned models, no bespoke queue infrastructure. Every one of those was considered. Every one was set aside for the version of the problem Quikest is actually solving.
Quikest has been in production since 2024. The product has been shipping continuously — persona-image regeneration UI in September 2025, documents and link-person UI in September 2025, login/register rebuild in October 2025, the current asset generation for this case study in July 2026. It is not a launched-and-done product; it is a working workspace being sharpened in the hands of the teams using it.
“Adoption metrics are held in-family. The architectural numbers — 3–5 personas per product, 4 models per pipeline, 3 export formats, 1 real-time cursor per teammate — are the ones on the page because they are the ones that describe the product.”
Persona generation is now a commodity — every tool ships some version of it. What teams consistently ask us about, and what they keep coming back for, is the Ask AI sidebar. The reference talking back is a different product than the reference sitting still.
We shipped the agent-context export in a single sprint. It has been the single most-quoted feature in demos to engineering-heavy teams. A small feature aimed at the underserved user in a category can outweigh the flagship feature everyone else already has.
The sage-and-gray palette, the single accent color, the refusal to add a purple gradient anywhere — those are marketing decisions as much as design ones. They tell buyers, before we say a word, that Quikest is not a hype product. The teams that respond to that are the teams we want to serve.
We do not have a “real-time collaboration” page in the product marketing because real-time is not a feature in Quikest — it is the material the product is made out of. The moment we would need to add such a page is the moment we would know the product had drifted.
The author-first AI writing surface built on radical transparency. Every AI edit logged, exportable as evidence.
Autonomous email negotiation agent for Red Dog Logistics. Wrong-order guard, margin engine, 100-point carrier scoring.
Eight-mode humanizer powered by an in-house fine-tuned model. Zero percent detected across six leading detectors.
If you’re planning a collaborative AI product where the shared reference matters — personas, briefs, decisions, anything a cross-functional team needs to hold in common — we would like to hear about it.
Quikest.ai · Live since 2024