The X Algorithm
Playbook

Predictions, not engagement exchange rates

Phoenix uses the viewer's engagement history to predict actions and values for candidate posts. Ranking combines weighted predictions. Retrieval, filtering and other processing also affect what is shown.

The README's August 14 clarification is explicit: weights scale predicted probabilities or continuous values, not raw engagement counts. A report does not cancel a fixed number of likes. Likewise, ten replies cannot be declared better than two hundred likes from those totals alone.

Positive and negative predictions both matter. Weight magnitudes alone do not tell a creator which action will dominate: prediction scales, the viewer, configuration and other stages matter too.

Thunder supplies in-network candidates; Phoenix retrieval and SimClusters supply out-of-network candidates. Being followed makes a post eligible for relevant retrieval, not guaranteed to be seen. Visibility filtering is distinct from ranking.

Five ways to make a post useful

These are editorial recommendations, not deductions of guaranteed outcomes from source code.

What this guide cannot promise

• No universal first-15-minute cutoff or fixed initial audience cluster follows from the public README.
• One or two posts per day is a possible starting cadence, not a hard algorithm limit. Test frequency against reader response and your capacity.
• Your own replies can build relationships, but do not establish control over the audience to which your future posts are recommended.
• Posting time depends on audience and context. Test your own time slots; there is no universal Eastern-time window.
• Do not infer that hashtags are unread or irrelevant from the absence of a particular hand-written feature. Content and engagement can both inform models.
• Do not promise that a smaller engaged account will always out-distribute a larger one.

Respect negative feedback: avoid spam, misleading claims and hostile exchanges. Predicted feedback and visibility rules can affect distribution, but a single action does not have a universal effect for every viewer.

Hook, value, evidence, invitation

Hook: State the problem or useful result.
Value: Explain one concrete idea.
Evidence: Show an example, source or measured outcome.
Invitation: Ask a relevant question or suggest a next step.

Use a thread only when the explanation needs it. Each post should carry useful information. There is no mandatory ten-post length or mathematically best final follow pitch.

Use that structure only when the numbers and result are yours to substantiate. A strong example is more useful than an invented performance claim.

Publish, learn, adjust

1. Review recent posts and choose one audience problem to address.
2. Gather a genuine example and supporting evidence.
3. Draft a concise post using the template above.
4. Publish at a time convenient for meaningful conversation; record the time and format.
5. Respond thoughtfully and note questions that recur.
6. Publish a follow-up that answers one of those questions.
7. Review available impressions, substantive replies, profile visits, follows and relevant conversions. Compare with similar posts before changing your approach.

Track multiple posts over time. Engagement totals are observations, not the model's predicted probabilities, and small samples cannot establish causality. Set goals from your own baseline rather than treating a fixed reply or follower target as an algorithm rule.