The X Algorithm
Playbook

How Your Posts Are Scored

Every post gets a score for every user who might see it. That score is computed by a Grok-based transformer model called Phoenix. The formula is literal:

Phoenix predicts the probability of 15 different actions the user might take after seeing your post. Each probability is multiplied by a weight — positive for good actions, negative for bad ones. The sum is your score.

This is the insight most creators miss: negative actions don't just reduce your reach — they actively drag your score down. A post that generates reports is worse than posting nothing.

The hierarchy that matters: follow > reply ≈ quote > share > like. Build every post around actions in that order, not the other way around.

The 5 Signals That Build Reach

What Will Bury Your Score

Most creators focus entirely on generating positive signals. The algorithm is symmetric — negative signals are weighted too, and they're not just the absence of positives. A post that generates reports is actively destroying your score.

How to Reach People Who Don't Follow You

The algorithm has two content sources. Thunder handles in-network posts — your followers already see you. The opportunity is Phoenix Retrieval: ML-based discovery that surfaces your posts to people who don't follow you yet, based on similarity between your content and their interests.

Phoenix uses a two-tower model: your content gets an embedding vector, each user has an interest embedding vector, and the algorithm matches them via similarity search. To reach a new audience, your embedding needs to be close to theirs.

What controls your content embedding

• Who engages with you — not who follows you. If ML engineers like your posts, the model learns you belong in the ML engineering cluster and distributes you there. If random accounts engage first, you drift into the wrong cluster.
• Your own outbound engagement — who you like, reply to, and quote shapes your own user tower embedding. Engaging with the community you want to reach positions you in that cluster before you publish.
• Vocabulary and structure of your posts — the transformer reads semantic meaning, not keywords. Writing in the language your target audience uses, about the topics they engage with, aligns your embedding to theirs.

The first 15 minutes are everything

The impression bloom filter means each post gets approximately one chance to appear in a given user's feed. Once shown, it won't be shown again. The first round of engagement determines which interest cluster seeds your distribution. Who engages first sets your audience trajectory for that post.

• Before publishing a major thread, engage in your target community — reply to 5–10 posts from well-followed accounts in your niche. This positions your user embedding in the right cluster.
• Line up 3–5 people who will reply or quote your best posts early. These signal seeds shape which users the algorithm shows your post to next.
• Post at 9–10 AM Eastern or 7–8 PM Eastern — when your target professional audience is actively scrolling and first-impression velocity is highest.

How to Structure High-Scoring Posts

The anatomy of a high-scoring post

[Hook]    — Creates a curiosity gap or makes a specific, contestable claim
[Body]    — 3–5 lines of dense value, each line independently shareable
[Proof]   — A number, outcome, or credential that signals authority
[CTA]     — A specific question or invitation to reply

Thread architecture

Post 1:   Tension hook — "The thing no one tells founders about X"
Posts 2–8: Value delivery — one insight per post, each self-contained
Post 9:   Synthesis or the most surprising conclusion
Post 10:  "If you got value from this, here's what to expect if you follow me: [clear promise]"

Post 10 is the most important post in the thread. It's the only one explicitly optimized for P(follow_author) — the highest-weight signal in the model.

The cadence rule

Phoenix applies a score decay to repeated author appearances in a single user's feed session. This is the author diversity attenuation — and it's the reason flooding doesn't work.

Replies and quote tweets to others don't trigger author diversity attenuation in your followers' feeds — but they actively shape your user tower embedding and generate profile click signals from the audiences of whoever you engage with.

What High-Scoring Posts Actually Look Like

The difference between a post that gets scrolled past and one that drives replies, quotes, and follows isn't talent — it's structure. These three rewrites show the exact pattern applied to common founder/creator mistakes.

Example 1 — The Vague Opener

Example 2 — The No-CTA Thread Ender

Example 3 — The Safe Take

The 7-Day Content Calendar

This schedule is designed around the author diversity attenuation rule (1–2 posts/day max), the cluster positioning principle (outbound engagement shapes your distribution), and the follow signal hierarchy (anchor threads drive the most follows).

Weekly signal targets

Your First 7 Days

This plan is designed to build your interest cluster, generate early engagement signals from the right audience, and launch your first algorithm-optimized thread — in that order.

The 5 Non-Negotiable Rules