helping developers get more out of @sanity_io \n (he/him) \n ask me about real-time text-fields fields in front of delightful JSON \n advisor @heavybit
Oakland, CA
- Tweets 23,566
- Following 2,639
- Followers 4,064
- Likes 15,877
knut retweeted
We wanted the agent on our Everything *[NYC] event site to do more than just answer event questions.
The site, blog, docs, and swag store are all separate Sanity projects. Each one feeds a Knowledge Base in Sanity Context, and the agent reads them all through a single Context MCP.
Kevin Green breaks it down in the full session. Watch it, then try Knowledge Bases (in beta). go.sanity.io/Rj27rKn
1
2
3
498
Is there any actual proof that "de-slopping" skills like this actually work? I have done the same exercise many-a-time for the last couple of years with limited success.
Slop is more than em-dashes (gone in opus 5.5 btw) and terms like "delve" and "seamless." AI-generated content is also recognized by document structure, how arguments are laid out, formatting (or the lack of; looking at you absent H3s), all the way down.
Sure, harder to spot in short copy, but you'll definitively spot it in longer prose.
LLM-generated content is an onion, not a banana.
And pouring a huge .md file into your context can't be good either (especially when the labs tell you to shave off skills and instructions for their capable models).
Introducing epstein.md 👋
A 927,400+ Character .MD File that's specifically trained to skip every vibe coded AI Slop design principle.
- no box shadows
- no emojis
- no '🟢 LIVE ' badges
- 0 blue/purple gradients
& religiously ignores 350 most common AI words (delve, vibe, seamless ...)
It's the result of 11 months of 8 hrs/day Twitter doomscrolling and constant additions to the codebase of our vibecoding ui.
Wholeheartedly hoping this finds the most design-needy indie hackers.
2
305
All I know is that a good chunk of these companies uses @sanity_io for their content; so their taste must be great.
7
391
Jev got 👀 📐🖼️ now!
Multimodal Jev and a new bounding box type from @typesafeai launched today as early access aka "dreamer mode" aka `typesafe_danger_zone`.
Took me about 10 minutes to integrate it with @sanity_io for automatically setting hotspot and crop based on your description.
And damn, this is actually useful, not just a fun demo.
How it works:
I have a Sanity Function that trigges on draft changes to only the `focus` and `image` fields (that way it works when agents work with content via mcp/api too).
Since Sanity has an image transformation pipeline, we can send downscaled optimized images to Jev without burning through the very cheap tokens too fast.
Then it asks Jev three things in one request:
→ Is it there? (yes/no)
→ Where is it? (a box)
→ Is everything around it plain background? (yes/no)
Jev's box comes back as fractions of the image. Sanity stores hotspot and crop as fractions of the image. So the integration is arithmetic, and code decides what to write, not the model. That is really cool.
Bonus: lint rules can look at images now. "Does the cover fit the post?" A photo of a pile of board games scores 3% on a post about agents. On a post about five best casual party games, the same photo gets 81% and passes. Fair.
Cost of the demo including experimentation: $0.04.
5
2
25
3,005
You can also read more about our early Jev-testing here—we're exploring multiple ways to use it for our product offering. More deep dives to come!
sanity.io/engineering/the-0-…
93
and demo time
Jev got 👀 📐🖼️ now!
Multimodal Jev and a new bounding box type from @typesafeai launched today as early access aka "dreamer mode" aka `typesafe_danger_zone`.
Took me about 10 minutes to integrate it with @sanity_io for automatically setting hotspot and crop based on your description.
And damn, this is actually useful, not just a fun demo.
How it works:
I have a Sanity Function that trigges on draft changes to only the `focus` and `image` fields (that way it works when agents work with content via mcp/api too).
Since Sanity has an image transformation pipeline, we can send downscaled optimized images to Jev without burning through the very cheap tokens too fast.
Then it asks Jev three things in one request:
→ Is it there? (yes/no)
→ Where is it? (a box)
→ Is everything around it plain background? (yes/no)
Jev's box comes back as fractions of the image. Sanity stores hotspot and crop as fractions of the image. So the integration is arithmetic, and code decides what to write, not the model. That is really cool.
Bonus: lint rules can look at images now. "Does the cover fit the post?" A photo of a pile of board games scores 3% on a post about agents. On a post about five best casual party games, the same photo gets 81% and passes. Fair.
Cost of the demo including experimentation: $0.04.
1
36
@typesafeai is enabling longer context windows (with a danger zone flag) up to ~262k - with the caveat that filling up longer context windows will degrade the performance.
17
the bounding box type is the first public example of @typesafeai's framing of "System One" models, it goes beyond just making decisions.
31
