A Daily AI Social Content Engine for a Beauty Salon

1. Trigger / content source — start on a schedule (e.g. daily 9 AM), manually, or from a new idea in Google Sheets, RSS, Airtable, Postgres, or a Meta webhook.
2. Post generation — a brand-agent LLM researches the topic live (Tavily/search) and writes a short, on-brand beauty-salon caption with a hook, tips, hashtags, and a soft CTA.
3. Visual prompt creation — a second agent converts the caption into a detailed photorealistic image prompt (subject, salon setting, lighting, mood) with no text or logos.
4. Image generation — gpt-image-1 (or a swappable provider) generates the matching visual, converted to a binary image file.
5. Archive — the image is uploaded to Google Drive.
6. Multi-channel publishing — the post and image are published to WordPress, Telegram, X, Facebook/Instagram, and LinkedIn, using only the channels that are configured.

• n8n — workflow orchestration (63 nodes, self-hosted)
• OpenAI gpt-image-1 — default AI image generation (swappable)
• Alternative image providers — Replicate, Google Imagen, Hugging Face, Kling, Runway, Leonardo, Ideogram, Clipdrop, NanoBanana, APITemplate.io
• LLM providers (interchangeable) — OpenAI, Anthropic Claude, Google Gemini, Mistral, DeepSeek, Groq, xAI Grok, AWS Bedrock, Azure OpenAI, Hugging Face, Ollama, OpenRouter
• Tavily / SearchApi — live web research grounding for captions
• Google Drive — generated-image archiving
• WordPress — CMS/blog publishing
• Telegram, X (Twitter), Facebook/Instagram, LinkedIn — social publishing
• Triggers — Schedule, Google Sheets, RSS, Airtable, Postgres, Meta/Facebook webhook
• n8n LangChain agents & buffer memory — text and visual-prompt generation with short-term memory

The brief: a salon's whole social presence, posted daily, with no daily effort.

The client wanted a hands-off engine that produces a polished, on-brand beauty-salon post every day — caption and matching image — and publishes it across all their channels at once, while staying simple enough to set up and flexible enough to swap models, providers, and platforms as needs change. We delivered this as a 63-node n8n workflow organised into five clean sections, built from the ground up to be modular: the language model, the image generator, the trigger, and the destination channels are all interchangeable.

Section 1 — Triggers and content sources. The engine can start in whatever way suits the salon's workflow. The default is a daily schedule (for example, 9 AM) for regular publishing, with a manual trigger for test runs. But it can also react to new ideas arriving from elsewhere: a Google Sheets row with a topic or reference link, a new RSS item (turning a fresh blog post or industry update into a social post), an Airtable record change, a Postgres event, or a Meta webhook. This makes the same engine work equally well for scheduled "evergreen" posting and for reactive, idea-driven content.

Section 2 — AI post generation. A "generate text" agent writes the post, acting as a brand agent for a modern beauty salon. Its system prompt is tuned to the audience (primarily women 20–45 who care about appearance, health, and relaxation) and to a warm, professional, educational tone covering hair, nails, skincare, make-up, and self-care. Critically, it always begins with live web research (via a Tavily / search tool) so facts and trends are current, and uses short-term memory for variety across days. Every post follows a proven structure: a strong hook, short readable paragraphs and mini-lists, evidence-based tips (no pseudoscience), minimal emojis, a soft non-aggressive offer, 3–6 relevant hashtags, and a simple call to action — all kept under 1024 characters so it's Telegram-ready.

Section 3 — Visual prompt creation. Rather than pairing the post with random stock imagery, a second "generate prompt" agent reads the finished caption and converts it into a detailed text-to-image prompt. It extracts the post's core message (which service, what result, what feeling), then specifies the model, the salon environment, lighting, mood, and composition in a photorealistic beauty-photography style — under hard constraints that forbid any text, numbers, or logos in the image.

Section 4 — AI image generation. That prompt drives image creation via gpt-image-1, producing a unique, on-brand visual that genuinely matches the post, which is then converted to a binary file for downstream use. The architecture is deliberately provider-agnostic: the image step can be swapped or duplicated to call Clipdrop, Ideogram, Replicate, Runway, Leonardo, Imagen, Hugging Face, Kling, or others, as long as they return an image file into the same outputs.

Section 5 — Distribution and posting. The finished post and image are fanned out to every channel the salon actually uses: uploaded to Google Drive for archiving, published to WordPress as a blog post, and posted to Telegram, X, Facebook/Instagram, and LinkedIn. Only the platforms that matter for the current strategy need to be configured; the rest stay inactive.

Why this design works. Its strength is flexibility without complexity. Setup is intentionally quick — connect your chosen chat model, your image provider, and the social accounts you use — yet every major component is interchangeable, so the salon can tune cost, style, and reach over time without rebuilding anything. Separating caption generation from visual prompt generation means the image always reflects the specific post rather than a generic theme, and grounding every caption in live research keeps content current and credible. The result is a genuinely set-and-forget engine: a fresh, on-brand, well-illustrated post goes out daily across the salon's entire social footprint, with the owner free to simply edit a system prompt if they want to shift voice, language, or focus.