Agent Blueprint: Content Repurposing Workflow

Use this when you regularly publish long-form content and want to systematically generate social distribution assets from each piece. Instead of manually writing X posts, LinkedIn posts, and newsletter blurbs for every article, this workflow produces them in one pass.

The blueprint is designed for n8n but the system prompts work in any automation tool (Make, Zapier, custom scripts) — only the workflow wiring changes.

Separating extraction from generation is the key design decision. Node 1 extracts the structured summary. Nodes 2-4 generate from that summary. This means each generation prompt receives a clean, consistent input rather than a raw article that varies in length and structure. Consistent input produces consistent output.

Each platform gets its own system prompt. A common mistake in content repurposing is using one prompt that says "adapt for X, LinkedIn, and newsletter." This produces three outputs that feel like the same thing reformatted, rather than three outputs native to their platforms. Separate prompts with platform-specific rules produce genuinely different content.

Running Nodes 2-4 in parallel saves time. They are independent of each other (all depend only on Node 1's output). In n8n, use a split node after Node 1 to run them concurrently. This cuts total execution time from ~30 seconds to ~12 seconds.

JSON output from every node enables clean data passing. Rather than parsing free text between nodes, each prompt outputs structured JSON that the next node can consume reliably. This prevents the cascading formatting errors that plague multi-step workflows.

Three X post variants give you options. Rather than generating one post and hoping it works, you get three angles. Pick the best one, or use all three across different days.

Add your voice configuration. Load the Editorial Voice skill (or your own style guide) as additional context in Nodes 2-4 for more consistent brand voice across all outputs.

Change the delivery node. Node 6 is modular. Swap it for whatever fits your workflow: post directly to Buffer or Typefully, save to a Notion database for review, send to a Slack channel for team approval, or push to a Google Sheet for batch scheduling.

Add a trigger. The workflow can be triggered manually, by webhook (when a new article is published), or by RSS feed (monitoring your blog for new posts). In n8n, the RSS trigger node works well for this.

Add more platforms. Duplicate the pattern for any additional platform: add a Node 5 with a Threads-specific prompt, or a Node 6 for YouTube community posts. Each platform gets its own prompt with platform-specific rules.

Adjust the X post character limit. If you use X Premium and want longer posts, change "Maximum 280 characters" to your preferred limit. The prompt structure works at any length.

The quality is bounded by the article quality. A thin, generic article will produce thin, generic social posts. The workflow amplifies signal — it does not create signal from nothing.

JSON output parsing can occasionally fail if the LLM produces malformed JSON. In n8n, add an error handling branch after each LLM node that catches JSON parse errors and retries once. Alternatively, add "Output valid JSON only. No markdown code fences, no explanation, just the JSON object." to each prompt.

This workflow does not handle images. Social posts often perform better with visuals. You would need to add a separate step for image selection or generation.

Claude: Reliable JSON output. Strong at maintaining distinct voice across platform-specific prompts. Recommended for the generation nodes (2-4).

GPT: Also reliable for this workflow. May occasionally wrap JSON in markdown code fences even when instructed not to. Add JSON parsing with fallback stripping of code fences in n8n.

For the extraction node (Node 1): Either model works. Claude tends to produce more concise key_points. GPT may include more context per point.

Cost optimisation: Use a smaller, faster model (Claude Haiku, GPT-4o-mini) for Node 1 (extraction is straightforward). Use a larger model (Claude Sonnet/Opus, GPT-4o) for Nodes 2-4 where quality matters more.