Courses
free + premium
Syntax
podcast
About
me
Blog
it's good
Tips
π₯ Real Hot
JavaScript
Notes
Speaking
and training IRL
/uses
Font?! Theme!?
Contact
me
Post not found
π
@wesbos
Tweets
42 minutes
0
7
2,243
Someone asked why I would use tagging instead of vector embedding + clustering I just tested it but find it hard with data that is faceted. Apple products grouped together, but keyboards and monitors aren't. I think next I'll try cluster the tags themselves into "areas"?
4 hours
10
155
19,600
No no no. You donβt get it. You pay $25 to have the same tool you had write the code to find the issues.
5 hours
1
12
3,293
Also getting icons for every single product was a few step skill: 1. "Enrich" the product, but asking an LLM to explain what it is and provide the URL from web search. Save to DB. 2. Ask the LLM to find an icon for that product - searching though 4-5 resources for the icon, and falling back to a favicon if none found 3. Write a script to figure out if the icon was black/white/color. Invert color if black. Almost every one is correct because it understands the context of the site. "basecamp" could be anything, but it nailed it.
6 hours
3
47
9,534
Here is a spot where I replaced a script with a skill. I had ~1100 tags and ~20,000 items in those tags for the new uses dot tech site. The data was dirty "MX 3" vs "mx3" inconsistent "Apple Macbook pro" vs "Macbook" and just roughly related "Mirrorless" vs "DSLR" We previously had a HUGE script that would try de-dupe common ones, but it's impossible. A skill that would go through tag + product worked amazingly well. I gave it access to some CRUD CLIs to merge + update data. After running it a few times I'm down to 180 tags and 11,000 products
Syntax Podcast: #985
March 9th, 2026
Stop putting secrets in .env
Listen Now β
Someone asked why I would use tagging instead of vector embedding + clustering
I just tested it but find it hard with data that is faceted. Apple products grouped together, but keyboards and monitors aren't.
I think next I'll try cluster the tags themselves into "areas"? 
Also getting icons for every single product was a few step skill:
1. "Enrich" the product, but asking an LLM to explain what it is and provide the URL from web search. Save to DB.
2. Ask the LLM to find an icon for that product - searching though 4-5 resources for the icon, and falling back to a favicon if none found
3. Write a script to figure out if the icon was black/white/color. Invert color if black.
Almost every one is correct because it understands the context of the site. "basecamp" could be anything, but it nailed it.
Here is a spot where I replaced a script with a skill.
I had ~1100 tags and ~20,000 items in those tags for the new uses dot tech site.
The data was dirty "MX 3" vs "mx3"
inconsistent "Apple Macbook pro" vs "Macbook"
and just roughly related "Mirrorless" vs "DSLR"
We previously had a HUGE script that would try de-dupe common ones, but it's impossible.
A skill that would go through tag + product worked amazingly well.
I gave it access to some CRUD CLIs to merge + update data. After running it a few times I'm down to 180 tags and 11,000 products