Content Crew

Inspiration

Most content creation tools force you to either use a rigid repurposing pipeline with no creative control, or build your own AI workflow on a node canvas that your actual marketing team will never touch. We wanted to build something different: an AI marketing team you direct, not a workflow you configure. The inspiration was simple — what if your AI tools felt like a real team showing you their work and reasoning, instead of a black box that outputs content?

What it does

ContentCrew is an AI-powered marketing operating system organized around five role-based AI Personas — Estratega, Creadores, Post-producción, Publicista, and Analista. You upload an existing video, and the team takes it through a 16-stage production pipeline: transcription, clip proposal, editing, subtitle burn-in, platform adaptation, thumbnail generation, SEO metadata, scheduling, and auto-publication to YouTube, Instagram, and TikTok.

What makes it different is how you interact with it. Every stage exposes its reasoning in plain language, anchored to the artifact it produced. An internal Slack-like layer shows agent-to-agent coordination in real time, and every agent message links directly to the artifact it's talking about — no floating text. Instead of approve/reject, users get six creative direction mechanisms: style references, structured controls, trend curation, visible edit-based learning, comparative branching, and manual takeover with fixed-point protection.

How I built it

The full stack runs on AWS: Next.js on Amplify Hosting, Aurora Serverless v2 PostgreSQL via RDS Data API (no VPC required), and an in-process orchestration engine that writes stage_runs, versions artifacts, and enforces a full dependency graph — so editing an upstream artifact automatically marks downstream stages as outdated. AI generation runs through Amazon Bedrock (Nova Pro as workhorse, Nova Lite for cheap transforms), with Amazon Transcribe for subtitle extraction, AWS Elemental MediaConvert for clip rendering and aspect-ratio adaptation, Nova Canvas for thumbnails, Nova Reel for premium video generation (credit-gated), and Amazon Polly for persona voices. Auth is Clerk. YouTube publishing is fully wired with real OAuth and resumable uploads.

A key architectural decision was keeping every external dependency behind a transparent local fallback — PGlite for the database, a deterministic local AI for generation — so the entire product is testable end-to-end with zero cloud infrastructure, and the cloud paths activate automatically when credentials are present.

Challenges I ran into

The hardest part was designing what "radical transparency" actually means at the UI level without turning the product into a technical dashboard. Exposing reasoning in natural language, anchored to concrete data and specific artifacts, required rethinking how agent messages, stage outputs, and creative direction panels relate to each other. They ended up as two views of one operation: the chat announces, the panel executes.

The async media pipeline — Nova Reel, MediaConvert, Transcribe — each has its own job lifecycle and failure modes. Unifying them behind a consistent two-phase pattern (submit → poll → artifact snapshot) while keeping the dependency graph coherent was a significant engineering challenge. The MediaConvert job template for mux + subtitle burn-in + aspect-ratio adaptation in a single job also took considerable iteration.

Accomplishments that I'm proud of

The artifact versioning system with is_manual protection (mechanism F) works exactly as designed: manual contributions are treated as fixed points the pipeline builds around and can never be silently overwritten by regeneration. The dependency graph that invalidates downstream stages when upstream artifacts change — and surfaces actionable "needs regeneration" badges to the user — is solid and battle-tested. And the dual learning loop keeping edit feedback and performance feedback structurally separate, with user-controlled scope and full reversibility, is the kind of detail that usually gets collapsed into a single opaque "the AI is learning" indicator.

What I learned

Repurposing is the right economic foundation for a product like this. Full video generation is genuinely expensive and belongs behind a credit gate; the interesting product lives in transcription, clipping, adaptation, and metadata — the ~$0.80/video path that's viable without venture funding. The other major lesson: building every external integration to its real boundary (implemented, config-gated, documented) rather than mocking it is the only way to know what you actually have. The LIMITATIONS.md file ended up being as important a document as the spec.

What's next for Content Crew

Completing the remaining platform integrations (Instagram Graph API and TikTok post-audit public posting), wiring live metrics sync for the Analista's 7d/30d reports, and implementing mechanisms C and E (active trend curation and comparative branching) as the next major creative direction layer. Longer term: the Persona marketplace and multi-client agency collaboration, where the internal Slack layer and per-campaign chat history become the audit trail clients actually want to see.

Built With

• amazon-web-services
• next.js