Published: 2026-04-14

Claude Routines Replace N8N: Build Automations in Natural Language

Chapters / key moments (click to jump — plays here on the page)

Nick Saraev makes the case — with live demos — that Claude Routines are a direct 1-to-1 replacement for N8N and other no-code automation platforms. Same input triggers (schedule, API call, webhook), same outputs (Slack, CRM, email), but built entirely in natural language. He demos email triage, a Fireflies transcript-to-proposal webhook, and migrating an existing N8N workflow by pasting its JSON into Claude Code.

Source video

"Claude Routines Just Dropped, And It's Perfect" by Nick Saraev — Watch on YouTube →

Key Takeaways

Routines are Claude's 1-to-1 replacement for N8N's middle layer. The old pattern: event → N8N drag-and-drop node chain → output platform. The new pattern: event → Routine natural language description → output platform. The trigger types and output destinations are identical; the middle layer is replaced by a prompt.
To migrate an N8N workflow: In N8N, hold Shift and drag to select all nodes, then Cmd+C to copy as JSON. In Claude Code, paste the JSON and say "use the routine generator skill to turn this N8N workflow into a routine." Done in seconds.
Webhook triggers enable event-driven automations. Nick demos connecting Fireflies (a meeting transcript service) so that when a transcript is ready, it fires a webhook → Routine generates a complete sales proposal automatically → proposal delivered to Slack. No cron, no polling — fires instantly when the transcript is ready.
Routines keep the full agentic framework. N8N reduces workflows to stateless compute chains — you lose the "agent" layer. Routines run the full W+A+T stack (Workflow + Agent + Tools together), which means Claude can self-correct mid-run, handle unexpected inputs, and compound on its own reasoning. This is fundamentally different from node-based automation.
Prompt discipline is different for routines than for skills. In a skill, you can steer mid-run. Routines are fully autonomous — Claude must succeed on the first try. Be narrower in scope, more explicit about steps, and include all context upfront. Think of it as writing a complete SOP, not a starter prompt.

Three Live Demos

Demo 1: Daily mailbox summary + draft. Routine checks Gmail unreads each morning, looks up prior email history with each sender for context, drafts reply emails, and sends a Slack summary. Nick runs it live via the "Run Now" button to show it working. Output: two email drafts plus a Slack message before he wakes up.

Demo 2: Transcript-to-proposal via webhook. When Fireflies finishes a meeting transcript, it fires a webhook to trigger the routine. The routine reads the transcript from the API payload, generates a complete business proposal using the agent's company context, and delivers it to Slack. Nick fires this by sending a curl request from Claude Code with the transcript as the payload body.

Demo 3: N8N workflow migration. Nick copies a Reddit scraper workflow from N8N as JSON, pastes it into Claude Code with the routine generator skill, and says "convert this to a routine, then also update it to send the output to Slack." Claude creates the routine and modifies it in one pass — what would have taken multiple N8N node edits happens in natural language in seconds.

Writing Good Routine Prompts

Nick's recommendations for routine prompts that actually work:

Write it like a complete SOP — steps 1, 2, 3 — not like a conversational request. The agent has no opportunity to ask clarifying questions.
Specify exactly what "done" looks like. If the output should be a Slack message, say so explicitly. If the agent should stop at a certain point or on a certain condition, make that explicit.
Include all credentials and connector references in the prompt — don't assume the agent will find them. If you're using a Gmail connector, say "use the provided Gmail connector."
Start with a narrow scope and verify it works with Run Now before scheduling. Expand scope incrementally.

Related on OpenClawDatabase

Claude Code Routines Setup — Nate Herk's guide to the technical setup, API keys, and trigger configuration
Claude Cowork Setup Guide — getting started with Claude Code and Cowork environments
OpenClaw Guide — alternative agent runtime with heartbeat scheduling for always-on workflows

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Three ways to fire a routine
Schedule, incoming webhook, or an API call. A routine can be triggered by incoming data or by outgoing data, and you can attach multiple triggers to the same routine. There is a run-now button that shows inputs and outputs live, which is the fastest way to test before scheduling.
Set the schedule from the trigger panel
Open the routine, change the trigger from API to schedule, and pick the cadence from the visual selector. The demonstrated case is an email triage flow set for 5:10am so the result is waiting before the user wakes at 5:20. Scheduled routines then appear on a calendar view.
Always end the prompt with a definition of done that produces an artifact you will see
The worked example ends with "once you're finished, use the Slack connector to send me an update" — because the conversation thread itself is not where you will look. Drafts landed in Gmail, and the summary plus the drafted replies arrived as a Slack DM. Without that final delivery step you have an automation whose output you have to go hunting for.
The replacement argument, stated precisely
The old shape was: an event feeds a no-code platform, you assemble a chain of drag-and-drop nodes with credentials, authentication and field mapping, and the output is pushed to Slack or a CRM. The new shape has the same event sources and the same destinations — <strong>routines replace the middle</strong>, substituting natural-language instructions for the node graph. That middle layer is where most of the build time went.
Chain to other agents through managed sessions
A routine can hand work to another managed agent rather than doing everything itself — each in its own siloed container for security. The demo pipes a meeting transcript in by API call and has a second agent generate the proposal.

Gotchas

Because the run is unattended and in the cloud, you cannot steer it. Instructions must be clearer and more complete than what you would type interactively — that is the real cost of moving an automation off your machine.
The presenter notes that the official documentation's examples skew heavily technical (backlog maintenance, alert triage, bespoke code review) in a way that obscures the general-purpose automation use case most people actually want.
Do not hardcode an API key into a demo prompt the way the video does — the presenter leaks his own on camera and says so.

← Back to News digest · See also: Claude Cowork guide