TL;DR
- Check templates first — Lindy has out-of-the-box knowledge base templates I should have started with
- Real workflows break — Google connections drop, folder IDs get lost, agents hang waiting for input
- Multiple approaches exist — agent-built workflow vs. off-the-shelf template, different tradeoffs
- Token efficiency matters — different workflow structures consume credits differently
- Use Claude to generate test data — I had it create 10 fake sales documents for the proof of concept
What I've Learned
This is Part 2 — building the actual Lindy workflow from Claude's spec. And debugging it live when things break.
The first lesson: check templates first. Lindy has out-of-the-box knowledge base templates. I jumped straight into the agent builder and built something custom — but I could have started with the template and modified it.
Real workflows break. The agent hung because it was waiting for me to reconnect Google Drive — no error message, just spinning. Folder IDs got lost. Slack channels didn't refresh. This is what real work looks like. Debug in public.
There are multiple valid approaches. The agent-built workflow and the off-the-shelf template both work. Different tradeoffs: the custom build gives more control, the template is faster to deploy. Token efficiency varies too — different structures consume different amounts of credits.
One useful trick: I had Claude generate 10 fake sales documents to test the workflow. Case studies, proposal templates, one-pagers. Beats manually creating test data.
Tools & Resources
- Lindy AI — agent builder and knowledge base templates
- Google Drive integration — for syncing sales documents
- Slack integration — for team queries
- Test data pattern — use Claude to generate fake documents for proof of concept