AI Automations

📖 3 min read resourcesautomationsworkflows Key Info

Copy-deploy automation recipes that put an LLM in the loop — inbox triage, meeting notes, digests, flashcards — using n8n or Zapier.

On this page 6

OverviewThe pattern behind every recipeYou’ll needThe reusable “AI transform” node (n8n)The recipesCost & safety notes

Practical, copy-deploy automations that drop an LLM into a real workflow. Each recipe is a small pipeline you can build in n8n (open-source, self-host or cloud) or Zapier (hosted, no-code) in 15–30 minutes.

These are blueprints, not one-click installs — you’ll wire your own credentials. But every recipe gives you the exact node sequence and the exact prompt to paste, so there’s no guesswork.

The pattern behind every recipe

Almost every useful AI automation is the same four-step shape:

flowchart LR
    A[Trigger<br/>new email / schedule / file] --> B[Fetch<br/>get the data]
    B --> C[LLM transform<br/>classify / extract / summarize]
    C --> D[Action<br/>label / create task / send]

The LLM step is just an HTTP request (or a native AI node) that sends a prompt and parses the JSON it returns. Master that one node and every recipe below is a variation.

You’ll need

• An LLM API key — Anthropic, OpenAI, or Groq. For high-volume automations, use a cheap, fast model: Claude Haiku 4.5, Gemini 3.5 Flash, or DeepSeek V4 Flash (see the Models Decision Guide and Cost Calculator).
• An n8n instance (cloud trial or npx n8n) or a Zapier account.
• API keys for whatever you’re connecting (Gmail, Slack, Notion, etc.).

The reusable “AI transform” node (n8n)

Every recipe calls an LLM via an HTTP Request node. Here’s the core config — set it once, reuse it everywhere:

{
  "method": "POST",
  "url": "https://api.anthropic.com/v1/messages",
  "headers": {
    "x-api-key": "={{ $env.ANTHROPIC_API_KEY }}",
    "anthropic-version": "2023-06-01",
    "content-type": "application/json"
  },
  "body": {
    "model": "claude-haiku-4-5-20251001",
    "max_tokens": 1024,
    "system": "<paste the recipe's system prompt here>",
    "messages": [
      { "role": "user", "content": "={{ $json.input }}" }
    ]
  }
}

The model replies in content[0].text. Ask it for JSON in the system prompt so the next node can parse fields directly. (Swap the URL/model for OpenAI or Groq — both are OpenAI-compatible at /v1/chat/completions.)

Tip — make output parseable. End every system prompt with: “Respond with ONLY valid JSON, no prose.” Then add an n8n Code node (JSON.parse($json.content[0].text)) or Zapier Formatter to turn it into fields.

The recipes

Recipe	Trigger	What the LLM does
Inbox Triage	New email	Classify priority + summarize + suggest a label
Meeting Notes → Action Items	New transcript	Extract owners, tasks, due dates, decisions
RSS → Daily Digest	Schedule (daily)	Summarize + group new items into one digest
Doc → Flashcards	New document	Generate Q&A flashcards (Anki-ready)

Cost & safety notes

• Cost: these run on per-token API pricing with no daily cap. At Haiku/Flash rates a triaged email or summarized article costs a fraction of a cent — but a runaway loop can add up. Cap max_tokens, and test on a few items before enabling the live trigger.
• Don’t auto-send. For anything outward-facing (replies, posts), have the automation draft and a human approve. Use the LLM to triage and prepare, not to act unsupervised.
• Privacy: these pipelines send your data to a model provider. For sensitive content, use a self-hosted model (Ollama + Llama/DeepSeek) behind the same HTTP node.