X Algorithm Optimizer

Optimize X posts using verified data from the open-source algorithm repositories:

• twitter/the-algorithm (2023, Scala)
• xai-org/x-algorithm (2025, Rust/Python, Grok-based)

Verified Engagement Weights

From the Heavy Ranker source code (relative scoring):

Signal	Weight	Interpretation
Reply engaged by author	75.0	Author replies to your reply
Reply	13.5	You reply to the post
Profile click	12.0	Visit author's profile
Conversation click	11.0	Click into thread
Conversation dwell (2+ min)	10.0	Read thread deeply
Retweet	1.0	Baseline amplification
Like	0.5	Passive endorsement
Video 50% completion	0.005	Watch half the video
Negative feedback	-74.0	"Show less often" / mute
Report	-369.0	Report the post

Key insight: One reply-with-author-engagement = 150 likes = 75 retweets.

Architecture: Thunder + Phoenix

Thunder (In-Network)

• In-memory post store consuming Kafka events
• Sub-millisecond retrieval of posts from accounts you follow
• ~50% of For You timeline comes from in-network

Phoenix (Out-of-Network)

• Retrieval: Two-tower model matches user embeddings to content embeddings
• Ranking: Grok-based transformer predicts 15 engagement types
• Enables viral distribution to non-followers
• ~50% of For You timeline comes from out-of-network

Post Optimization Checklist

1. Hook (First 50 characters)

The transformer model uses early stopping patterns. Optimize for scroll-stop:

• Start with contrarian/unexpected statement
• Use specific numbers over vague claims
• Remove hedging words ("I think", "maybe", "perhaps")

2. Trigger Replies (Weight: 13.5-75x)

Replies dominate the scoring formula. End posts with:

• Direct question ("What's your take?")
• Debate frame ("Change my mind")
• Experience prompt ("Drop yours below")
• Unfinished list ("What's #3?")

3. Enable Reposts (Weight: 1x, but distribution multiplier)

Include 1-2 standalone quotable lines that work when shared without context.

4. Avoid Negative Signals (Weight: -74 to -369)

• Strong opinions invite engagement
• Offensive content triggers blocks/reports
• One report erases ~369 likes in scoring

5. Timing

Post when followers are online to maximize initial velocity. Phoenix activation requires strong early signals.

What the Reddit Post Got Wrong

Common misinformation circulates about X algorithm "weights" using a 1-10 scale. The actual weights from source code are:

• NOT "Reposts 8/10, Replies 7/10, Likes 6/10"
• ACTUAL: Reposts 1.0, Replies 13.5, Likes 0.5 (replies are 27x more valuable than reposts)

The "19 signals" categorization and specific "velocity threshold" formulas are not in the open-source code.

Premium/Verified Boost

From source code: Blue Verified accounts receive a 2-4x multiplier on their final score, applied after engagement weights.

References

For detailed technical architecture, see references/architecture.md.