A while back I worked through a round of X / Twitter's public recommendation code and wrote it up as an article. Later I found this material works better as a tool you can ask at any time: when you hit a specific question, it looks up the matching code and the conditions it applies under.
So I made For You Skills, for Claude Code, Codex, Cursor, and other tools that support Skills.
Project: https://github.com/alexliyu7352/for-you-skills
If you also post on X, you may have been stuck on questions like these:
- I already posted once today. Should another one wait until tomorrow?
- For a long piece, should it go out as a regular Post, a Thread, or an Article?
- People say "a second post loses 40% of its reach." Where does that claim actually come from?
- My draft is done. What is concretely wrong with its opening, its length, and its images?
You can ask it all of these directly.
What I care about is whether the basis of an answer can be checked. So besides the research findings, I also put the matching original code links, the conditions each finding applies under, and the unresolved questions into the knowledge base. When you ask, it looks up the relevant entries first and then gives a recommendation; where the public evidence is not enough, it says so.
Take the "second post loses 40%" case. The point there is to pin down what the code's "second" is actually counting, and whether a score multiplier can be read directly as an impressions ratio. The demo keeps that Q&A and the matching source code, so you can check it yourself.
If you have a draft, you can also have it reviewed. It checks the actual opening, measures the length, considers the first image and the ones after it, judges which publishing format fits the content, and gives specific edits. There are 8 skills in total, and one entry point selects the workflow by question.
Install with this one command for now:
npx skills add alexliyu7352/for-you-skills --skill x-assistant
Once it is installed, try asking in your tool:
Use x-assistant: I posted on X this morning, and I have another post on a different topic for this afternoon. Should I publish it today or tomorrow? Give me a recommendation, and explain what the relevant source code supports and what it cannot prove.
The boundaries: the knowledge base is based on research into one specific version of the public source (snapshot 5ed7963, research cutoff 2026-10-29), so it cannot confirm the full live configuration, and it cannot guarantee followers or reach. Model, web access, image generation, and similar capabilities are provided by the tool you use.
The 78-second case demo and the full record are here: https://github.com/alexliyu7352/for-you-skills/blob/main/docs/demo/README.md, and the original research article is still here: https://blog.go5.si/posts/x-algorithm-teardown/
I am looking for people who already use AI tools and also need to post on X to try it. A question you have actually been stuck on, or a draft you are about to publish, makes the best test.
I hope you can help test it and see whether it helps. If anything does not work well for you, please send me feedback, and I will keep improving it.
Top comments (0)