A few months back, I spent a weekend building a carousel workflow in Lindy.
The idea: take a podcast transcript, run it through a Lindy agent, and automatically generate a series of formatted slides. What used to take me a couple of hours of manual design work… done in minutes. I was pretty happy with it.
Then I went to update the workflow. New models were out. Gemini 3 and ChatGPT 5.1 had just dropped. Figured it was a good time to upgrade.
Within 20 minutes, I was confused.
The intro section wasn’t showing up. The layout I’d dialed in over weeks was completely wrong. I kept tweaking prompts, trying to rephrase things, adjusting the instructions. Nothing worked. Both of the “smarter” models were just… not following my format.
Eventually, I gave up and went back to Gemini 2.5 Flash. The cheapest model available.
First try. Perfect output.
Why Newer Models Sometimes Follow Instructions Worse
This seems counterintuitive at first. If a model is smarter, shouldn’t it do a better job?
Here’s the thing: “smarter” often means the model has been trained to be more helpful, more creative, more interpretive. It fills in gaps. It tries to anticipate what you probably meant. It improvises.
For a lot of tasks, that’s great. You want the model to help you think. To push back. To synthesize.
But for a very specific formatting task — like generating a carousel with a particular structure, every time, exactly right — that interpretive intelligence works against you. The model decides your instructions might be too rigid. It tries to make it better. It changes things.
Gemini Flash was trained differently. It reads the instructions and does them. That’s it.
I’ve started thinking of it like hiring for different roles. If you need someone to process invoices to a very specific format, you don’t hire the most creative person in your office. You hire the person who reads the instructions carefully and follows them exactly. Different jobs need different people. And different tasks need different models.
The Automation That’s Been Running for 10 Years
Here’s a related thing I’ve been thinking about.
In my Gmail, I have a filter that automatically labels any email containing words like “receipt,” “order,” or “thank you for your purchase” and drops them into a folder called Purchases.
I set that up over 10 years ago.
It still runs every single day. I basically forgot it existed until someone asked me how I keep my inbox clean, and I started listing all the automations I have. That filter showed up on the list and I thought, “oh right… that thing.”
That’s what a good automation does. It disappears into the background and just works. You stop thinking about it.
The mistake people make is chasing new tools and new models before the old ones have even finished paying off. The Gmail filter cost me 10 minutes to set up a decade ago. I have no idea how many hours it’s saved me since.
A Pattern Worth Noticing
I’ve been using Lindy for a while now, and I keep noticing that my relationship with it feels similar to something that happened about 15 years ago with OmniFocus.
When Asian Efficiency was just getting started, I built a strong early relationship with the OmniGroup team. They were building something that felt ahead of its time — a serious task manager for serious people. We partnered on content, wrote deep tutorials, and helped their community understand how to actually use it. OmniFocus became one of the most-read topic areas on the site for years.
Lindy feels like that. It’s a small, focused company building something real. The product keeps getting better. The team is responsive. And the people who are learning to use it now — really use it, not just dabble — are going to have a significant advantage in the next few years.
The principle here isn’t really about Lindy specifically. It’s about betting early on things that are working, even if they’re not mainstream yet. And sticking with the version that actually works for you, even when newer options arrive.
How to Think About AI Model Selection
Here’s a simple framework I’ve landed on:
When to use a smarter (more expensive) model:
- Research and synthesis
- Writing first drafts
- Analyzing complex data
- Conversations that need nuance
When to use a cheaper, more compliant model:
- Following a specific output format
- Structured data extraction
- Repeatable automation steps
- Anything that needs to work exactly the same every time
The test is simple: does this task need intelligence or compliance? If it needs intelligence, use the best model you have access to. If it needs compliance, the cheaper model probably does it better.
And if an AI tool isn’t doing what you want — before you spend an hour rewriting prompts — try swapping the model. You might solve it in 5 minutes.
One More Thing
The automation you set up 10 years ago might still be the most valuable one you have.
Check your Gmail filters. Check your old Zapier or Make workflows. Check the Shortcuts you built on your phone in 2019 and forgot about. Some of those are probably still running quietly, doing exactly what you told them to do.
That’s what good automation looks like. You stop thinking about it. It just works.
If you want to dig into AI model selection and workflow design in more depth, that’s exactly what we cover in our AI workshops. The next one is May 31 in Fort Worth — details at asianefficiency.com.
