For years I’ve been working on concepts and prototypes of predecessors of Siftler. The premise is simple: instead of an algorithm that learns what keeps you scrolling, what if you had one that learns what your most trusted sources find worth reading?
That’s collaborative filtering applied to information feeds. Feeds in the system or on the network are compared to your benchmark feed, which contains links to information that you have appreciated. Links that appear in sources most similar to your benchmark rise to the top. No profiling. No engagement optimisation. Just the logic of shared attention. That algorithmic logic has been fermenting in my head since I first saw book recommendations on Amazon.
I’ve been working on a non-AI, open-source implementation in a few prototyping projects, including most recently with the South-Eastern Finland University of Applied Sciences.
The challenge was always implementation. Building a deterministic, auditable piece of software takes time and resources I haven’t had. For one thing, I’m not a software programmer.
Now I figured out a different approach: I wrote the algorithm as an AI prompt.
The prompt instructs an AI application — and I’ve tested this first on Anthropic’s Claude — to execute Siftler’s collaborative filtering logic step by step. You submit sets of links, it compares them, and it returns a ranked RSS feed of recommendations. The calculations are fully transparent. The prompt itself is plain text that you can read, audit, and modify.
Today I’m releasing it publicly as version 0.1.
You can find it here: https://codeberg.org/josschuurmans/siftler-ai-prompt
It’s licensed under AGPL-3.0, which means you’re free to use and modify it, but any derivative work must remain open under the same terms.
This is not a finished product. It runs in Claude, it produces results, and it has real limitations. Technical ones like bandwidth, token use, and the inherent non-determinism of AI, but also broader ones: valid societal concerns about AI’s risks and its energy footprint.
My goal is to demonstrate the concept clearly enough that developers, democracy advocates, and potential partners can see what we’re building toward: a transparent, deterministic, non-AI implementation that can one day power alternative feeds on social platforms, run on the decentralised web, and serve organisations that need tailored information monitoring.
My previous post on decoupling the algorithm from the platform addresses aspects of that vision.
If you try Siftler’s AI prompt, I’d love to hear what you think. File an issue on Codeberg, or get in touch directly.
Follow the project at https://josschuurmans.com/en/the-information-stack
Reach me at jos.schuurmans@cluetail.com.
