Generate LinkedIn carousel images from text with Mistral AI & S3 Storage

AI Carousel Caption & Template Editor Workflow Overview This workflow is a caption-only carousel text generator built in n8n. It turns any raw LinkedIn post or text input into 3 short, slide-ready title + subtext captions and renders those captions onto image templates. Output is a single aggregated response with markdown image embeds and download links. --- Workflow Structure - Input: Chat UI trigger accepts text and optional template selection. - Core AI: Agent cleans input and returns structured JSON with 3 caption pairs. - Template Rendering: Edit Image nodes render title and subtext on chosen templates. - Storage: Rendered images uploaded to S3. - Aggregate Output: Aggregate node builds final markdown response with embeds and download links. --- Chat Trigger (Frontend) - Trigger: When chat message received - UI accepts plain text post. - allowFileUploads optional for template images. - SessionId preserved for context. AI Agent (Core) - Node name: AI Agent - Model: Mistral Cloud Chat Model (mistral-small-latest) - Behavior: - Clean input (remove stray formatting like \n and but keep emojis). - Produce exactly one JSON object with fields: postclean, title1, subtext1, title2, subtext2, title3, subtext3. - Titles must be short (max 5 words). Subtext 1 or 2 short sentences, max 7 words per line if possible. - Agent must return valid JSON to be parsed by the Structured Output Parser. Structured Output Parser - Node name: Structured Output Parser - Validates agent JSON and prevents downstream errors. If parsing fails, stop and surface error. Normalize Title Nodes - Nodes: normalize title,name 1, normalize title,name 2, normalize title,name 3 (and optional 4) - Map parsed output into node fields: title, subtext, safeName (safe filename for exports). Template Images - Source: Google Drive template PNGs (download via Google Drive nodes) or provided upload. - Keep templates high resolution and consistent aspect ratio. Edit Image Nodes (Render Captions) - Nodes: Edit Image 1, Edit Image2, Edit Image3, Edit Image3 (or Edit Image3/Edit Image4 as available) - MultiStep operations render: - Title text (font, size, position) - Subtext (font, size, position) - This is where caption text is added to the template. Upload to S3 - Nodes: S3 - Upload rendered images to bucketname using safeName filenames. Confirm public access or use signed URLs. Get S3 URLs and Aggregate - Nodes: get s3 url image 1, get s3 url image 2, get s3 url image 3, get s3 url image 4 - Merge + Aggregate: Merge1 and Aggregate collect image items. - Output Format: output format builds a single markdown message: - Inline image embeds !image(<s3-url>) - Download links per image. --- Integrations Used | Service | Purpose | Credential | |---------|---------|-----------| | Mistral Cloud | AI agent model | Mistral account | | Google Drive | Template image storage | Google Drive account | | S3 | Store rendered images and serve links | Supabase account | | n8n Core | Flow control, parsing, image editing | Native | --- Agent System Prompt Summary > You are a data formatter and banner caption creator. Clean the user input (remove stray newlines and markup but keep emojis). Return a single JSON object with postclean, title1/subtext1, title2/subtext2, title3/subtext3. Titles must be short (max 5 words). Subtext should be 1 to 2 short sentences, useful and value adding. Respond only with JSON. --- Key Features - Caption only output: 3 short slide-ready caption pairs. - Structured JSON output enforced by a parser for reliability. - Renders captions onto image templates using Edit Image nodes. - Uploads images to S3 and returns markdown embeds plus download links. - Template editable: swap Google Drive background templates or upload your own. - Zero guess formatting: agent must produce parseable JSON to avoid downstream failures. --- Summary A compact n8n workflow that converts raw LinkedIn text into a caption-only carousel with rendered images. It enforces tight caption rules,