Paid Generation Fan-Out

Paid Generation Fan-Out is my operational standard for maximizing the utility of every paid LLM API call, ensuring a single prompt generates tailored content for a minimum of five distinct distribution channels.

What it is

This approach dictates that any interaction with a paid LLM API must result in content suitable for at least five different distribution surfaces. The core principle is that single-use generation is a design failure. Instead of asking an LLM for a single tweet, I ask it for a structured JSON object containing a tweet, a LinkedIn post, an email subject, a short blog summary, and an image prompt—all from one request. This isn't about simply repurposing existing content; it's about architecting the initial generation to be multi-faceted and purpose-built for diverse channels from the outset. It ensures that the cost of the LLM call is amortized across multiple valuable outputs, making each token spent more efficient.

Why I do it this way

I run a multi-product studio with AI as the team. Every resource, especially paid API calls, must deliver maximum return. The cost of a Claude API call, for instance, is the same whether it generates one sentence or a complex, multi-part structured output. By fanning out the generation, I drastically reduce the per-unit cost of content production. This directly supports a profit-first operating model. It also ensures that when a new piece of content or product update is ready, its promotional and informational assets are simultaneously generated and queued for distribution across social, email, and web. This eliminates manual content creation bottlenecks and keeps all channels fresh without additional human effort. For example, when a user purchases a guide on Pregnancy Power Hour, the system immediately generates the guide, the follow-up emails, and the social promotion, all from one LLM interaction.

How it works in practice

The process begins with a trigger—often a Stripe webhook indicating a product purchase or a scheduled event. This trigger hits a Vercel function, which then hands off to VERA, my custom agent orchestration layer. VERA crafts a single, highly detailed prompt, sent to Claude API or Gemini, requesting a structured JSON output. This JSON schema defines fields for a blog post, a social media post (X and LinkedIn), an email body, a short blurb for a PDF, and an image generation prompt. Once the LLM returns this structured data, VERA's sub-agents take over.

For instance, the "Social Agent" extracts the social post content, formats it for specific platforms, and queues it for publication. The "Email Agent" uses the email body and subject line, then dispatches it via Resend. A "PDF Agent" takes the PDF blurb and integrates it into a pre-defined React component template, rendering it to a final PDF using the principles outlined in PDFs as React Components. Concurrently, an "Image Agent" leverages the image prompt to generate visuals via Imagen, which are then attached to social posts or embedded in web content, integrating with my approach to Imagen over Stock Photos. This entire chain, from trigger to multi-channel distribution, is automated. For Total Formula 1, when a new jargon term is added, this system generates the glossary entry, a social explainer, an email snippet, and a relevant image, all at once.

Where this breaks down

While efficient, this approach introduces complexity. The initial prompt engineering for a fan-out system is significantly more demanding than for single-purpose calls. Getting an LLM to consistently produce five distinct, high-quality content formats within a single structured output requires extensive iteration and refinement of the prompt templates and output schemas. Quality control is another critical area; without robust Critic Gates for Mass Content in place, the quality of individual fan-out pieces can degrade. Each output needs to be fit for its specific purpose and audience. Lastly, maintenance overhead exists. Changes to platform APIs (e.g., X character limits, LinkedIn content policies) or internal brand guidelines require updates across the fan-out logic, which can be time-consuming to manage across multiple agents and output formats.