Build a Self-Improving TikTok Carousel Workflow in n8n: Gemini Plans, fal.ai Renders, OpenAI Writes, Auto-Publishes

Input: product info + historical data sheet; Process: daily analytics refresh → Gemini generates strategy → fal.ai renders 4× 9:16 images → OpenAI writes caption → Upload-Post publishes; Output: published TikTok carousel + updated skill log

Who

Content operators or developers with an n8n instance and experience configuring API credentials

You get

Automatically generate and publish a 4-slide TikTok carousel daily, iteratively improving strategy based on historical performance data

You need

n8n · google-sheets · fal.ai · upload-post

Steps

1. Trigger and Configuration

   Set up Daily Schedule (every 24 hours) or Manual Trigger in n8n. In the Set Config node, enter: product_name, product_description, upload_post_user, upload_post_api_key, spreadsheet_id. All downstream nodes read from this single configuration source.

2. Refresh Historical Post Analytics

   Read Performance Log reads the Google Sheets Performance Log table. Filter Posts for Stats (Code node) skips rows with complete data or older than 7 days, keeping only request_ids needing updates. Get Post Analytics calls the Upload-Post API (/api/uploadposts/post-analytics/{request_id}?platform=tiktok) for views/likes/comments/shares. Extract Stats (Code node) calculates engagement_rate. Update Stats writes back to the sheet.

3. Build Learning Context

   Aggregate compiles updated data. Get Profile Analytics calls the Upload-Post API (/api/analytics/{user}?platforms=tiktok) for followers, impressions, etc. Read Agent Skill fetches the skill memory. Read Fresh Data re-reads Performance Log. Prepare Context (Code node) compiles product info, skill memory, profile data, and post history into agent_skill / profile_analytics / performance_summary objects.

4. Strategy Agent Plans the Carousel

   Strategy Agent (AI Agent node) uses Gemini 2.5 Pro (models/gemini-3.1-pro-preview) + Think Tool + Output Parser (Structured Output). The systemMessage defines the role as Larry (TikTok growth strategist) and requires strictly valid JSON output: title, strategy_notes, prompt1-4, hashtags, engagement_hook. Set Vars flattens output fields into workflow variables.

   Execution prompt · Source text。Strategy Agent node systemMessage field (escape JSON before pasting)

   Copy prompt

   You are Larry, an elite TikTok growth strategist with persistent memory.
   
   === YOUR ACCUMULATED SKILL (learnings from ALL past runs) ===
   {{ $('Prepare Context').first().json.agent_skill }}
   === END SKILL ===
   
   Create 4-image TikTok carousel slideshows (9:16 vertical) to promote a SaaS product.
   
   CRITICAL - Each image MUST contain visible designed TEXT:
   - Slide 1: Powerful HOOK text (big, bold, scroll-stopping)
   - Slide 2: Key feature/benefit with text overlay
   - Slide 3: Another angle/use case/proof with text
   - Slide 4: Strong CTA text
   
   Apply your skill learnings. Build on winners. Avoid losers.
   Output strictly valid compact JSON only.

   Execution prompt · Source text。Strategy Agent node text field (user prompt)

   Copy prompt

   === PRODUCT ===
   {{ $('Prepare Context').first().json.product_name }}: {{ $('Prepare Context').first().json.product_description }}
   
   === TIKTOK PROFILE ===
   {{ $('Prepare Context').first().json.profile_analytics }}
   
   === POST HISTORY ===
   {{ $('Prepare Context').first().json.performance_summary }}
   
   === TASK ===
   Create an optimized TikTok carousel using your skill + this data.
   
   Return ONLY valid JSON:
   {
     "title": "<TikTok caption with hook, max 150 chars>",
     "strategy_notes": "<What and WHY based on learnings, max 200 chars>",
     "prompt1": "<Slide 1: HOOK - visual + bold text to display, 9:16>",
     "prompt2": "<Slide 2: FEATURE - visual + text about product, 9:16>",
     "prompt3": "<Slide 3: BENEFIT - visual + text, 9:16>",
     "prompt4": "<Slide 4: CTA - visual + call-to-action text, 9:16>",
     "hashtags": "<8-12 hashtags>",
     "engagement_hook": "<Question for comments>"
   }

5. Render Four Carousel Slides

   Generate Image 1-4 are four parallel HTTP Request nodes calling fal.ai (https://fal.run/fal-ai/nano-banana-2), each POSTing prompt1-4 with num_images:1, aspect_ratio:'9:16', output_format:'png'. Each followed by Extract URL 1-4 (Split Out nodes) extracting the url from the images array. Merge Images combines four branches; Aggregate Images aggregates into a single item.

6. Caption, Publish, and Log

   Generate Description (OpenAI node, gpt-4.1) generates a TikTok caption (≤300 chars) from product_name, title, hashtags, engagement_hook. Upload to TikTok (uploadPost node) publishes the photo carousel with four image URLs, title, caption, and tiktokAutoAddMusic enabled. Log to Sheets appends post_date, title, description, strategy_notes, prompts_summary to Performance Log.

   Execution prompt · Source text。Generate Description node messages[0].content (system role)

   Copy prompt

   You write viral TikTok carousel descriptions. Short, punchy, emojis ok.

   Execution prompt · Source text。Generate Description node messages[1].content (user role)

   Copy prompt

   =Write TikTok description for {{ $('Set Config').first().json.product_name }}.
   Hook: {{ $('Set Vars').first().json.title }}
   Hashtags: {{ $('Set Vars').first().json.hashtags }}
   Question: {{ $('Set Vars').first().json.engagement_hook }}
   
   1.Hook 2.Value 3.CTA 4.Hashtags 5.Question
   Max 300 chars. ONLY the description.

7. Self-Improve Agent Skill

   Update Skill (OpenAI node, gpt-4.1) reads current skill, latest post data, full history, and profile analytics, then outputs an updated skill file (≤2000 chars) with sections: What Works / What Doesn't Work / Key Metrics / Strategy Playbook / Product Angles. Write Skill writes skill_content to the Agent Skill table using id='skill' as match key via appendOrUpdate.

   Execution prompt · Source text。Update Skill node messages[0].content (system role)

   Copy prompt

   You maintain a concise skill file for an AI content creator. Analyze data, write actionable learnings.

   Execution prompt · Source text。Update Skill node messages[1].content (user role)

   Copy prompt

   === CURRENT SKILL ===
   {{ $('Prepare Context').first().json.agent_skill }}
   
   === LATEST POST ===
   Title: {{ $('Set Vars').first().json.title }}
   Strategy: {{ $('Set Vars').first().json.strategy_notes }}
   
   === ALL DATA ===
   {{ $('Prepare Context').first().json.performance_summary }}
   Profile: {{ $('Prepare Context').first().json.profile_analytics }}
   
   Write an UPDATED skill file (max 2000 chars). Sections:
   ## What Works (data-backed)
   ## What Doesn't Work
   ## Key Metrics (averages, best post, trend)
   ## Strategy Playbook (top 3 to try next)
   ## Product Angles (best ways to promote)
   
   Be specific with numbers. Update existing insights, don't just append. Output ONLY the skill content.

After importing this template into n8n, complete setup in this order: 1) Copy the Google Sheet template and replace spreadsheet_id; 2) Configure credentials for Google Sheets OAuth2, Gemini, OpenAI, and fal.ai (HTTP Header Auth: Authorization: Key <FAL_KEY>); 3) Sign up for Upload-Post and enter username + API key; 4) Update product_name and product_description in Set Config to match your product; 5) Run manually first to verify each node output; 6) Check that Agent Skill writes correctly. To customize, edit the product positioning in Strategy Agent's systemMessage and text, or swap the fal.ai model.