I recently revisited a Make.com scenario I built a few months back-before beta agents were even a thing. Originally inspired by a setup from Jack Roberts (our top-tier coffee-fueled broki), this was my early attempt to build an agent-like image analysis flow using Make, Telegram, and Anthropic’s Claude 4.
I tweaked the system back then just for fun, but I had a reason to dive back in: someone asked if it’s possible to retrieve and analyze images from Telegram using Make. That gave me the perfect excuse to bring this old build back to life-and once again test how solid Claude is when it comes to deep image understanding.
Testing AI Image Smarts the Right Way
We’re not talking about basic OCR or detecting objects. I’m talking real comprehension-feeding in a complex diagram and getting a structured, smart breakdown of what it means.
For this test, I uploaded a tech-heavy invoice processing diagram via Telegram. Important side note: Telegram compresses image uploads and gives you three versions. From what I’ve seen:
-
The first is super compressed.
-
The second is mid-range.
-
The third one is usually the original, and that’s the one you want for proper analysis.
So I filtered the system to grab that third image and sent it to Claude via Make.
What Happened Next?
Claude (Sonnet, in this case) nailed it.
It didn’t just say “this is a diagram” or “there’s some invoices.” It correctly recognized it as an invoice processing system, then broke down relationships, identified entities, and returned a structured analysis-almost like a smart JSON response.
It caught things like:
-
Invoices having multiple transactions
-
Status changes in workflow
-
The general structure and intent of the system
All from one image. No manual hinting. No step-by-step breakdowns. Just pure visual reasoning.
Why Claude Still Wins
There’s been chatter about new models coming for the throne-GPT-4o, Gemini, you name it. But I haven’t seen any of them consistently match Claude when it comes to actual deep image understanding.
And to be clear: I didn’t invent the idea of breaking down visuals like this. I’ve seen others share incredible ways to analyze images and return structured outputs. I just built on top of it, added some workflow fun, and made sure the Telegram integration works.
Want to Try It?
Here's the prompt I used so you can experiment yourself. Play around. Push it. Let me know if you think anything else out there is doing a better job.
But as of now? Claude’s still the king of AI vision.
- Max
