Clario
The Problem
Producing long content is expensive and most of it is used once. A single video, podcast or article then has to be turned into summaries, posts, threads and newsletters by hand, which takes hours and rarely sounds like the person or brand publishing it.
Why Existing Approaches Fall Short
- •Generic AI tools write text that does not sound like the person or brand publishing it.
- •Most tools solve one job, so summaries, rewrites and drafts end up scattered across different products.
- •Pasting content into a chat window loses context, history and any consistent structure between sessions.
What Was Built
- ✓Summaries in ten formats, including executive brief, action items, SWOT, key quotes, sentiment and a plain language explanation.
- ✓A YouTube summariser that accepts standard, short and live links and falls back across the transcripts available for a video.
- ✓A remix studio that turns one piece of content into ten platform specific formats at the same time.
- ✓An AI assistant that keeps conversation history and applies the selected brand voice to every answer.
- ✓A brand voice library where writing samples are stored once and applied across summarising, remixing and chat.
- ✓A dashboard and content calendar showing what was created and how much of the monthly allowance has been used.
- ✓Accounts with email or Google sign in, monthly usage limits and subscription billing.
How It Works
Add the source material
Paste text, or drop in a YouTube link to pull the transcript automatically.
Choose the outcome
Pick a summary format, a remix target, or start a conversation with the assistant.
Review and adjust
Every result is editable, and a single format can be regenerated without losing the others.
Take it away
Copy the result or export it as plain text or Markdown for wherever it will be published.
Key Architectural Decisions
Separate instructions per output format
Each format has its own prompt, length limits and structure rules. That produces drafts people can actually use instead of generic text, and it is the main reason output quality holds up.
One shared usage system
Every AI action is recorded against a single monthly allowance, so limits, plan changes and billing stay predictable for the user and for the code that gates features.
Access rules enforced in the database
Row level security means an account can only ever read its own rows, even if a query is written incorrectly later. The rule lives with the data instead of in the interface.
AI calls stay on the server
Keys never reach the browser, requests can be rate limited per user, and the AI provider can be swapped without touching the interface.
Stack & Rationale
Next.js App Router
Server rendering for fast first loads and API routes in the same codebase.
TypeScript
Catches mistakes while building rather than after launch.
Supabase
Database, authentication and row level security in one place.
Groq running Llama models
Fast, low cost text generation for a product that makes many requests.
Stripe
Subscriptions, invoices and a customer billing portal without building payments from scratch.
Recharts
Usage charts without a heavy charting dependency.
Tailwind CSS and shadcn/ui
A consistent interface that stays quick to maintain.
Vercel
Deployment, preview builds and serverless functions with no server to look after.
Lessons & Outcomes
- •Prompt quality per format matters more than which model is used. Generic instructions produce generic output.
- •Usage limits have to be designed together with billing from the start, not bolted on afterwards.
- •People edit AI output far more than expected, so editing and regeneration deserve to be first class features.