This is my own tool, not a client project. The workflows belong to the public Zie619/n8n-workflows collection, not to me. Full write-up and code: github.com/Melwis880/n8n-organizer.
The problem
When scoping an automation project, the useful question is “what has already been built for this?” A public library of n8n workflows answers it, but only if you can search it. The collection has 2,057 JSON files in 187 folders. Raw n8n JSON is long, full of node parameters and canvas positions, and some files are copies of others. Loaded as-is into a notebook tool like NotebookLM, it means hundreds of sources and a lot of noise.
The goal: a few large, clean sources, each workflow described the same way, with a category you can trust and see the reason for.
How it works
- Scan the folder and skip what is not a workflow, recording why.
- Measure each workflow: working nodes, real connections, triggers, branches, services.
- Score it into one of three categories (
Data_Integration,AI_Content,Orchestration_Reliability). Every point is written out as a reason line. - Remove duplicates, including copies that differ only in ids, positions or node order.
- Write one Markdown file per category, plus a summary and a trace of every decision.
No LLM, no network. Node parameters, credentials and URLs never reach the output, because real workflows often hold API keys in their parameters.
The finding: the newest data was broken
Before trusting any numbers, I compared node types across the collection’s git history. Two later
commits had damaged it: one added thousands of disconnected error nodes, another turned AI nodes
into empty noOp nodes while keeping their names (320 nodes still named “OpenAI Chat Model”).
A classifier that reads node types cannot be right on that data, so the tool runs on the last
clean version (ae8cf6dc).
The result
| Measure | Value |
|---|---|
| Workflows analysed | 2,047 |
| Unique workflows written | 1,973 |
| Duplicates removed | 74 |
| Errors | 0 |
| Run time | about 7 seconds |
| Output | 3 files, each under the NotebookLM word limit |
Accuracy, checked by hand on independent samples of 24 workflows: 16 of 24 before tuning, 24 of 24 (with a scoring signal) after. Samples of 24 are small; these are checks, not a benchmark.
Quality work: 123 tests over 57 written scenarios. To test the tests, 167 deliberate bugs were planted one at a time and 153 were caught. Three security reviews before publishing fixed issues like crafted file names breaking a run and a hash that could leak values from workflow parameters.
Takeaways
- Look at the data before tuning the model. Here the newest data was the problem.
- An honest “no signal” beats a confident default.
- Hand-checked samples, kept separate from the tuning sample, are what make accuracy numbers honest.