X For You Algorithm Optimization Skill
Overview
This skill provides deep insights into X's For You feed algorithm (based on the xai-org/x-algorithm repository) to help create content optimized for maximum reach and engagement. It translates technical ML mechanics into actionable content strategy.
The Algorithm in Simple Terms
Two-Stage System
-
Retrieval (Two-Tower Model): Finds relevant posts from millions of candidates
- User Tower: Encodes your engagement history into a "user embedding"
- Candidate Tower: Encodes posts into "post embeddings"
- Similarity Search: Dot product matches users with content
-
Ranking (Grok Transformer with Candidate Isolation): Scores and orders candidates
- Takes user context (engagement history) + candidate posts
- Candidates CANNOT attend to each other (independent scoring)
- Predicts probabilities for 17+ engagement actions
Key Design Principles
- No hand-engineered features: Pure ML, the model learns everything from engagement data
- Candidate isolation: Ensures scores are consistent and cacheable
- Multi-action prediction: Predicts likelihood of each engagement type separately
- Hash-based embeddings: Scalable for billions of posts
Engagement Types & Strategic Implications
The model predicts probabilities for:
Positive Signals (Increase Score)
- favorite (like) - baseline positive
- reply - HIGH VALUE: conversation starter
- retweet - amplification, distribution
- quote - HIGH VALUE: adds commentary + amplification
- click - link clicks, media expansion
- profile_click - author exploration
- video_view (vqv) - video engagement
- photo_expand - image interaction
- share - sharing to DMs, etc.
- dwell - time spent reading
- follow_author - follows the poster
Negative Signals (Decrease Score)
- not_interested - user marked "Not Interested"
- block_author - severe negative
- mute_author - negative
- report - severe negative
Continuous Actions
- dwell_time - actual seconds spent viewing
Scoring Mechanics
Final Score = Σ(weight_i × P(action_i)) + diversity_adjustment - negative_signals
Weighted Combination
Each engagement action has a learned weight in the final score calculation:
- Reply, Quote, Retweet, Share likely have positive weights
- Follow author likely positive
- Not interested, block, mute, report have NEGATIVE weights
Implication: Content that predicts high probability of NEGATIVE actions gets pushed DOWN.
Author Diversity Adjustment
After weighted scoring, author diversity adjustment:
- First post from author: 100% of score
- Second consecutive post: multiplied by decay_factor (e.g., 0.5)
- Third: decay_factor² (e.g., 0.25)
- Floor ensures minimum visibility (e.g., 0.1)
Implication: Don't spam the feed with multiple posts in a row from same author.
Video Quality View (VQV) Weight
Video posts get an additional weight IF video duration > threshold (e.g., > 2 seconds)
Implication: Short videos may not benefit from video-specific weighting.
Filters & Quality Gates
Pre-Scoring Filters (Remove before ML scoring)
- Duplicates
- Age > threshold (e.g., 24-48 hours old)
- Self-posts (your own posts)
- Repost deduplication
- Ineligible subscription content
- Previously seen posts
- Previously served in session
- Muted keywords
- Blocked/muted authors
- Core data hydration failures
Post-Selection Filters (Final validation)
- VF Filter: deleted, spam, violence, gore, etc.
- Dedup conversation: avoid multiple branches of same thread
Implication: Fresh, unique, non-spammy content only. Old posts get filtered out.
Content Strategy: Key Insights
1. DESIGN FOR HIGH-VALUE ENGAGEMENTS
The algorithm predicts engagement probabilities. To maximize score:
Trigger REPLIES
- Ask questions
- Create controversy (within brand safety)
- Pose "hot takes" that beg counterpoints
- Use "fill in the blank" formats
- Polls (built-in X feature)
- Tag relevant accounts to encourage response
Trigger QUOTES
- Create highly shareable "quote tweet bait"
- Provocative statements with clear positioning
- Data points that merit commentary
- Bold predictions or strong opinions
- "Thread 🧵" hooks that make quoting natural
Trigger RETWEETS
- Pure value: "this everyone should know"
- Emotional resonance (awe, anger, joy)
- Social proof: "everyone's talking about this"
- Helpful resources/threads
- Alignment with audience identity
Trigger VIDEO VIEWS
- Use native video (not embedded from YouTube)
- First 3 seconds MUST hook (autoplay preview)
- Vertical format for mobile
- Captions for silent viewing
- Duration: long enough for meaningful view (> threshold) but not too long
Trigger PHOTO EXPANDS
- Multi-image carousels encourage tapping
- "Swipe to see X" or "Slide 3 is the 🔑"
- High-quality, intriguing images that make users want to zoom
Trigger DWELL TIME
- Dense, thoughtful content that causes reading pause
- Long threads that keep people scrolling
- Text in images that requires reading
- Medium-length posts (not too short, not too long)
Trigger PROFILE CLICKS
- Mention interesting accounts
- Have a compelling bio/profile yourself
- Tag brands/people that others would explore
Trigger FOLLOWS
- Post content that makes people think "I need more of this"
- Consistent niche positioning
- Social proof (follower count, verification)
- Value-add tweets, not just promotion
2. AVOID NEGATIVE SIGNALS LIKE THE PLAGUE
Negative actions HURT your future reach:
Prevent NOT INTERESTED
- Stay on-topic for your audience's interests
- Avoid clickbait that doesn't deliver
- Don't post irrelevant content
- Test content before scaling (use metrics)
Prevent BLOCKS/MUTES
- Don't be spammy
- Don't post offensive content
- Don't @mention people unnecessarily
- Don't engage in fights publicly
- Don't post NSFW unless your audience expects it
Prevent REPORTS
- Never mislead/fraud
- No hate speech
- No impersonation
- No spam/mass marketing
Watch Dwell Time: Very short dwell (<1s) might correlate with negative implicit signals (though not explicitly modeled).
3. UNDERSTAND CANDIDATE ISOLATION
Each post's score is computed independently in the same request batch.
Implication: The algorithm doesn't compare your post against others in real-time; it's an absolute score based on user+post match.
- Your post's fate is determined by its predicted engagement probability with that specific user, not relative ranking in the moment.
- However, author diversity applies AFTER scoring, affecting which posts actually get served.
4. AUTHOR DIVERSITY MATTERS
Even if your content scores high, multiple posts from same author get attenuated within the same feed response.
Implication:
- Don't post many times in quick succession. Your later posts that day get less reach even if better content.
- Space posts: 2-4 hours minimum between posts.
- If you have multiple great posts, schedule them apart.
5. RECENCY IS CRITICAL
Age filter removes posts older than threshold (likely 24-48 hours).
Implication:
- Only recent posts get shown.
- Timeliness matters: news, trends, hot takes.
- Fresher content advantage.
6. USER'S ENGAGEMENT HISTORY DRIVES RELEVANCE
The model ingests the user's recent engagement sequence (likes, replies, retweets, etc.) to determine what they'll like.
Implication:
- Target audience definition is everything.
- Content should align with what your followers already engage with.
- Consistency builds a clear engagement signature, improving reach to the right users.
- Branching into new topics reduces relevance until new engagement data accumulates.
7. HASH-BASED EMBEDDINGS = CONTEXTUAL SEMANTICS
Posts and authors are embedded via hash tables, not explicit categories.
Implication:
- The model understands semantic meaning, not just hashtags or explicit topics.
- Natural language processing determines topic affinity.
- Keywords matter, but context and language patterns matter more.
- onboard the "language" your audience uses.
8. OUT-OF-NETWORK (OON) DISCOVERY
Phoenix retrieval finds relevant posts from accounts the user doesn't follow.
Implication:
- You can reach new audiences even with low follower count.
- Quality content that predicts engagement can break out of your follower bubble.
- The two-tower model matches your content's "embedding" to users seeking that content.
- Use broad language that matches high-engagement patterns; don't over-niche to the point of zero discoverability.
9. CANDIDATE ISOLATION ENSURES FAIR PLAY
Candidates cannot attend to each other during ranking, so your post's score isn't affected by which other posts are in the candidate set.
Implication:
- Your content quality is absolute, not relative.
- A+ content always scores high regardless of competition.
- Conversely, mediocre content won't be "saved" by weak competition.
- Consistency of quality is key.
10. PRE-SCORING FILTERS ARE GATES
Many posts get filtered out before even reaching the ML model.
Implication:
- Posts must pass quality and eligibility checks.
- No duplicate content (even your own reposts).
- Freshness required.
- Not already seen by user (previously seen filter).
- No muted keywords (user-specific).
- Author not blocked/muted by user.
- Core data present (text, author info, etc.)
Content Optimization Checklist
Before Posting
For Video Content
For Threads
Posting Frequency
After Posting
How to Use This Skill
For Content Ideation
-
Query: "How do I make this post more likely to get replies?"
→ Answer: Frame it as a question, use controversial/discussion-worthy angle, tag relevant accounts.
-
Query: "Why did my video post underperform?"
→ Answer: Check if video duration > threshold, first 3 seconds hook, native upload, vertical format.
-
Query: "Should I post multiple times today?"
→ Answer: Author diversity decay means 2-4 hour spacing. High-quality spaced posts > spammy cluster.
For Strategy
-
Query: "What engagement type should I optimize for?"
→ Answer: Depends on goals. For reach: retweets/shares. For conversation: replies/quotes. For authority: dwell time/long threads.
-
Query: "How do I break out to new audiences?"
→ Answer: Optimize for broad appeal actions (retweet, share, video view). Ensure content has clear topic semantics for retrieval matching.
-
Query: "What's more important: likes or replies?"
→ Answer: Likely replies > likes (higher weight in scoring). But depends on model weights. Focus on engagement types that indicate deeper commitment.
For Troubleshooting
-
Query: "Why is my reach dropping?"
→ Check: Negative signals up? Post frequency too high? Audience mismatch? Content quality?
-
Query: "Why did this flop while my last post succeeded?"
→ Compare: Topic alignment, format, hook strength, negative signals triggered.
Technical Notes
Model Architecture
- Transformer: Grok-based, likely similar to Grok-1 architecture but adapted for recsys
- Embedding size: configurable (D dimension)
- Context length: user + history (S positions, e.g., 128) + candidates (C positions, e.g., 32)
- Attention mask: candidate isolation ensures independent scoring
Data Requirements (Inferred)
- User features: hashed user ID, following list, preferences
- History: sequence of engaged posts with:
- Post ID (hashed)
- Author ID (hashed)
- Actions taken (multi-hot vector)
- Product surface (mobile/web/API)
- Candidates: post + author features (similar to history but without actions)
Training Signal
- Positive engagements train the model to associate content with those actions
- Negative engagements (blocks, mutes, reports, not interested) train the model to avoid that content
Limitations & Unknowns
The public repository omits:
- Exact weight values (FAVORITE_WEIGHT, REPLY_WEIGHT, etc.): These are proprietary tuning parameters
- Exact thresholds: Age filter duration, video minimum duration, diversity decay factor, floor
- Exact model hyperparameters: embedding size, number of layers, attention heads
- Training data regime: How often models retrain, data sources
- Post-selection filter details: VF filter specifics
- OON scorer details: out-of-network content adjustment
- Real-time features: Any online features added at inference time beyond what's in the batch
Strategy implication: We must infer optimal practices from first principles and common patterns, not exact numbers.
Practical Experiment Framework
When testing content strategies:
- Control variables: Post similar content types, only change one factor
- Metrics (track per-post):
- Impressions (if available via analytics)
- Engagement rate (engagements / impressions)
- Breakdown: likes, replies, retweets, quotes, video views, photo expands, profile clicks, follows
- Negative signals: blocks, mutes, reports, not-interested (if accessible)
- Dwell time (if available)
- Estimated reach (impressions ÷ follower count)
- Statistical significance: Test multiple posts per variant (n=10+)
- Time control: Compare posts at similar times of day, similar follower activity patterns
- Audience consistency: Ensure same audience segments seeing variants
- Content quality parity: Hooks must be equally strong across variants
Actionable Takeaways (TL;DR)
- Design for replies and quotes: They likely carry more weight than likes.
- Use video natively: Optimize first 3 seconds, use vertical, include captions, ensure >2s duration.
- Space your posts: Minimum 2-4 hours between posts to avoid author diversity decay.
- Evoke dwell: Medium-length posts (200-500 chars) with dense value encourage reading time.
- Tag strategically: Encourage profile clicks and replies by mentioning relevant accounts.
- Avoid spammy patterns: Too many posts, clickbait, repetitive content trigger negative signals.
- Be consistent: The model learns user preference from your engagement history; niche down.
- Freshness matters: Timely content outperforms old evergreen posts.
- Format for expansion: Carousels and images that benefit from zoom/expand.
- Test to learn: The exact weights are unknown; experiment to see what moves your metrics.
Integration Example
A social media manager can use this skill to:
- Pre-screen post ideas through the "optimization checklist"
- Diagnose underperforming posts by checking which filters/scorers may have penalized them
- A/B test different hooks and formats with a principled approach
- Plan posting cadence to maximize author diversity benefits
- Choose format (video vs. image vs. text) based on desired engagement outcomes
Questions This Skill Can Answer
- "How does X's For You algorithm decide what to show?"
- "What makes a post go viral on X?"
- "Why am I not getting reach on my posts?"
- "Should I post threads or single tweets?"
- "What's better: video or images for reach?"
- "How often should I post per day?"
- "What's the ideal tweet length?"
- "Do hashtags help with the algorithm?"
- "How important are replies vs likes?"
- "Can I game the algorithm by asking for engagement?"
- "Does quoting vs retweeting matter?"
- "Should I reply to my own tweets to boost engagement?"
- "What time of day is best for posting?"
- "Do polls get more reach?"
- "How do negative signals affect my account?"
This skill distills engineering knowledge into marketing strategy. Use it wisely.
