Offroading with Claude and n8n

I tried to use Claude and n8n like a weekend offroad rig. Not for a polished product, just to see what breaks when you stop driving on the nicely paved “chat in a UI” road.

The idea: wire Claude into n8n to do a few ugly, real tasks that I actually hit during the week:

• Clean up notes I dump into a Google Doc
• Turn half-baked ideas into task lists in ClickUp/Notion
• Summarize support threads from Slack and log them somewhere

Nothing visionary. Just “I don’t want to manually copy-paste and re-word this anymore.”

What I actually set up

1. Google Docs → Claude → Notion
   Trigger when a doc gets tagged with a special string at the top, like [[PROCESS]]. n8n grabs the doc, sends it to Claude with a structured prompt, and posts back to Notion as a cleaned summary plus action items.

2. Slack → Claude → ClickUp
   Use a Slack reaction (✅) as the trigger. n8n pulls the thread, Claude turns the mess into a clear ticket (title, description, acceptance criteria), then n8n creates the ClickUp task.

3. RSS / URLs → Claude → “Should I care?”
   Feed URLs into n8n (RSS or manual). Claude returns: 3-sentence summary, bullet list of “why it matters to me,” and a yes/no/“later” flag. n8n files it into a Notion database filtered by that flag.

What worked

• Claude is good at reshaping messy language. Support threads from Slack turned into decent tickets with almost no tweaking. The model is pretty good at pulling intent out of long back-and-forths.
• n8n is solid for the plumbing. Once the nodes were wired (Slack → Claude → ClickUp), it just ran. The visual editor is fine for this kind of flow, and the built-in HTTP node covered gaps for tools that didn’t have ready-made integrations.
• Simple, repeated patterns shine. The best flows were: “when X happens, send all this text to Claude with a strict output format, then shove it into Y.” That’s it. The more I tried to be clever, the worse it got.
• JSON-in / JSON-out prompts worked best. Having Claude reply with a fixed JSON schema ({"title":"","summary":"","tasks":[...]}) made downstream mapping in n8n painless. Less glue code, fewer weird failures.

What didn’t work

• Overcomplicated flows fell apart. I tried a chain like: “classify the message, pick a ‘mode’, then call Claude again with a different prompt based on the mode.” Too many branches, too many chances for malformed output. Debugging those in n8n felt like crawling under the car in the mud.
• Reliance on perfect formatting is fragile. Even with clear instructions, sometimes Claude would:
• Wrap JSON in backticks
• Slip in an explanation line above the JSON
• Miss a field when the input was weird
• That breaks strict JSON parsing in n8n. I ended up adding an extra “cleanup” step or switching to a more forgiving parser. It worked, but it’s duct tape.
• Latency matters more than I thought. When you’re just chatting, 5–10 seconds is fine. As an automation step, waiting ~8 seconds every time I reacted to a Slack message got old. It didn’t kill the idea, but it changed which flows felt worth it. Low-frequency, high-friction tasks were okay; high-frequency things were not.
• Non-determinism is annoying in ops. Human-facing: creative variation is nice. Workflow-facing: I want boring. Sometimes Claude would rephrase ticket titles in a way that made search worse, or group tasks differently between similar inputs. Not wrong, just inconsistent.
• Error reporting is weak out of the box. When something failed (rate limits, malformed JSON, integration hiccups), it wasn’t always obvious whether the problem was my prompt, Claude’s output, or n8n’s node config. I had to add more logging nodes just to see what was going on.

What I’d keep using

• Slack → Claude → Tickets stays. It removed friction from creating “proper” tickets. Good enough is actually good enough here.
• Docs → Claude → Summaries is useful, but only for things I really don’t want to rewrite manually. I turned off auto-triggering; now I run it on demand so I don’t burn API calls on junk.
• News / URL triage was nice in theory, but in practice I just stopped caring. If I don’t want to skim the article, I usually don’t want a summary either. I paused that flow.

Would I do it again?

Yes, but I’d treat Claude+n8n as a utility, not a brain.

• Use it to compress and reformat text I already decided matters
• Keep flows short: trigger → Claude → one destination
• Force strict schemas and expect occasional cleanup
• Assume it will be a bit slow and a bit fuzzy

As an offroad rig, it did what I wanted: showed where things scrape. Some lines are now obviously “don’t bother automating this.” Others are boring enough that I don’t think about them anymore, which is kind of the point.