R — Original Text (Reading)
"We have eliminated every single hand-engineered feature and most heuristics. The Grok-based transformer does all the heavy lifting by understanding your engagement history. During transformer inference, candidates cannot attend to each other—only to the user context, making scores consistent and cacheable. Final Score = Σ (weight_i × P(action_i)). Positive actions (like, repost, share) have positive weights. Negative actions (block, mute, report) have negative weights, pushing down content the user would likely dislike."
— xAI, X Recommendation Algorithm.md · Key Design Decisions / Scoring and Ranking
Self-translation: We have eliminated all handcrafted features and most heuristic rules. The Grok transformer learns relevance on its own from your interaction history. During ranking, candidate posts cannot see each other, and only attend to user context, making scores stable and cacheable. Final score = weighted sum of the probabilities of each action, with positive actions (like/repost/share) having positive weight and negative actions (block/mute/report) having negative weight, actively suppressing content the user does not like.
I — Methodological Skeleton (Interpretation)
X's For You recommendation feed is driven by a pipeline: dual-source retrieval → hydration filtering → multi-action scoring → diversity → selection. Three core design decisions form the reasoning skeleton for understanding "why content is recommended":
- No Handcrafted Features: The Grok transformer learns relevance on its own from interaction history, rather than relying on manually defining "what content is good." Whether content is relevant is learned from user behavior patterns, not preset rules.
- Candidate Isolation: During scoring, each post is evaluated independently and cannot "see" other posts in the same batch, and only attends to user context. Your post is scored by its own predicted quality, not by being "better than other posts in the same batch"—quality is the only lever, not relative timing.
- Multi-Action Weighting: It does not predict a single "relevance" score, but instead predicts the probabilities of 15+ actions (like/reply/repost/bookmark/click/dwell/follow……), with positive actions having positive weight and negative actions (block/mute/report) having negative weight, and the weighted sum yields the final score.
2025 update: OON share has risen to 60-70%, and the algorithm relies more on content relevance and interaction quality than on account size, benefiting small and medium accounts; the downranking of main posts with links has been removed; rage-bait/negative content is actively suppressed. Implications for creators: a single post is scored by its own quality (not by follower count), so optimize for bookmarks/replies (high positive weight) rather than only chasing likes, and avoid triggering block/mute/report.
A1 — Application in the Book (Past Application)
Case 1: xAI open-sources the For You source architecture
- Problem: How does X's For You feed decide which posts are recommended to users?
- Use of the Method: xAI open-sourced the complete recommendation system source code (github.com/xai-org/x-algorithm), showing the full pipeline of dual-source retrieval (Thunder following feed + Phoenix global discovery) → Phoenix Grok transformer scoring → multi-action weighting → candidate isolation → author diversity → selection, and clearly labeling the three major design decisions (no handcrafted features/candidate isolation/multi-action prediction).
- Conclusion: The algorithm relies entirely on a transformer to learn relevance from interaction history, with no manual feature engineering; each post is scored independently and does not compete with posts in the same batch.
- Result: This lets creators understand "why it was recommended" from an engineering perspective—not black-box mysticism, but an analyzable model of "predicted interaction probability × weight."
Case 2: The impact of the 2025 algorithm update on small and medium accounts
- Problem: What does the 2025 algorithm update mean for small and medium accounts?
- Use of the methodology: AIP Tutorial 3.1.2 interpretation of the 2025 algorithm rules — OON share has shifted from early In-Network dominance to about 60-70%, and the algorithm relies more on content relevance and engagement quality than on account size; the downranking of main posts with links has been removed (favorable to knowledge-sharing accounts); rage-bait/negative content is actively downranked (because it triggers the negative weights of block/mute/report).
- Conclusion: A single post from a small or mid-sized account can enter the recommendation pool on its own quality without relying on follower base; optimizing saves/replies (high positive weight) is more effective than chasing likes only; avoid rage-bait (negative weight lowers the overall score).
- Result: The inference of candidate isolation was verified — posts are scored by their own quality, and the OON tilt gives small-account content more opportunities to be discovered globally.
A2 — Trigger Scenarios (Future Trigger)
In what situations will users need this skill?
- Can a small account get recommended: The user asks "I only have 200 followers, can my content get recommended? Will the algorithm only push big accounts?" — they want to understand the opportunities for small accounts in the algorithm.
- Why no traffic/no recommendation: The user asks "Why does my post get no exposure?" "My post sinks as soon as it's published, how does the algorithm judge it?" — they want to diagnose using the algorithm's logic.
- Wanting to understand the recommendation mechanism itself: The user asks "How does the For You algorithm work?" "What is X's recommendation mechanism?" — they want to understand the underlying logic.
- The impact of engagement types on the algorithm: The user asks "Which matters more to the algorithm, saves or likes?" "Why do controversial posts have high engagement but lower exposure instead?" — they want to understand the differences in weight among multiple actions.
- Downranking of rage-bait/negative content: The user asks "Why does posting controversial content actually cause traffic to drop?" — they want to understand the negative weight mechanism.
Language signals (these should activate it when they appear in the user's words)
- "For You algorithm" / "recommendation mechanism" / "why no traffic" / "can a small account take off" / "OON" / "algorithm downranking" / "which matters more, saves or likes"
- "For You algorithm" / "recommendation" / "why no impressions" / "algorithm penalty" / "how does the algorithm work"
Distinction from adjacent skills
- Difference from
x-longtail-strategy: longtail talks about "long-tail distribution mindset/increasing the probability of entering the top 10%" and is at thestrategic level; this skill talks about "how the algorithm scores/why something gets recommended" and is at themechanism level. If the user says "my post didn't blow up and I'm falling apart" → longtail; if the user says "why didn't the algorithm push me" → this skill. - Difference from
x-comment-engagement: comment-engagement talks about "how to engage to gain exposure" and is at theaction level; this skill talks about "how the algorithm evaluates the value of engagement" and is at themechanism level. If the user says "how to engage in the comments section" → comment-engagement; if the user says "the impact of saves and replies on algorithm weight" → this skill. - Difference from
x-cold-start-playbook: cold-start talks about "actively heating up in the first 24h to break through the initial pool" and is at theexecution level; this skill talks about "why the algorithm relies on initial engagement signals" and is at themechanism level. If the user says "how does a new account build volume" → cold-start; if the user says "how does the algorithm score" → this skill.
E — Executable Steps (Execution)
When the skill is activated, the agent should execute the following steps:
-
Confirm the type of user question
- Determine whether the user wants to "understand the mechanism" (→ explain the algorithm structure) or "diagnose why there is no traffic" (→ use the algorithm logic to infer possible causes).
- Completion criterion: Clarify whether the user's intent is "how the algorithm works" or "why my post was not recommended".
-
Analyze one by one using the three major design decisions
- Dual-source retrieval: Can the post enter OON (strong content relevance, can be discovered globally) or only In-Network (visible only to followers)? An OON share of 60-70% means small accounts also have a chance.
- Multi-action weighting: What actions did the post trigger? Likes but no bookmarks/replies (a single type of interaction)? Negative signals (block/mute/report)? Positive actions have positive weight, negative actions have negative weight—bookmarks/replies weigh more than likes, and negative signals actively suppress.
- Candidate isolation: Posts are scored independently based on their own quality and do not compete with the same batch—so content quality is the only lever, not the relative timing of "being better than the post above you".
- Completion criterion: Map the three design decisions one by one to the user's specific situation.
- Stop condition: If the user only wants to know "how the algorithm works" rather than diagnose, skip to step 3 and give the inference directly.
-
Give action inferences based on the algorithm logic
- If it is a small account: An OON share of 60-70% is favorable; a single post can enter the recommendation pool on quality alone, without relying on follower count.
- If interactions are of a single type: Optimize for bookmarks/replies (high positive weight) rather than only chasing likes; bookmarks are a "future action signal", and replies are "deep participation".
- If there is a rage-bait tendency: Avoid triggering block/mute/report (negative weight actively suppresses content, and the composite score actually decreases).
- If it includes links: In 2025, link demotion for main posts has been removed, so links can be freely integrated to improve usefulness.
- Completion criterion: Give specific action recommendations based on the algorithm logic, rather than a vague "post good content".
B — Boundary
Do not use this skill in the following situations
- Do not use it when purely writing content and not asking about the mechanism: "Help me write a tweet" and "How should this Hook be revised" are the responsibility of
x-short-content-craft, and this skill does not guide content writing. - Do not use it for mindset anxiety about "what if my post does not go viral": That is the long-tail distribution mindset of
x-longtail-strategy(accepting a median of 42 as normal, with the goal of increasing the probability of entering the top 10%); this skill discusses the mechanism, not mindset. - Do not use it for "how to interact to gain exposure": That is the comment interaction strategy of
x-comment-engagement; this skill only explains "how the algorithm evaluates the value of interactions", not "how to interact". - Do not use it for "how a new account gains traction/how to warm up in the first few hours": That is the cold-start execution of
x-cold-start-playbook; this skill only explains "why the algorithm relies on initial signals".
Failure modes the author warns about in the book
- ce22 — Rage-bait is actively demoted by the algorithm: Although rage-bait/negative content triggers replies (high P(reply)), it also triggers a large number of block/mute/report actions (higher negative weight), so the composite score actually decreases. In 2025, the algorithm clearly prefers positive/informational content. Posting controversial content in exchange for interactions is a trap of high short-term interaction and long-term suppression.
- ce21 — Game-the-algorithm tactics fail after the platform tightens: Mutual follows/comment-section sniping/matrixing are essentially exploiting current algorithm loopholes. Platform anti-cheat keeps evolving (in 2025, link downranking was abolished and rage-bait is being cracked down on); techniques that work today may be directly downranked tomorrow. The algorithm will change—don't exploit loopholes; content capability is the asset that can migrate across algorithm cycles.
The author's blind spots / era limitations
- Highly dependent on the 2025-2026 algorithm state (OON share 60-70%, link downranking abolished); the algorithm may change again at any time, and specific thresholds will drift—the OON share may be adjusted again in future versions.
- What xAI open-sourced is thearchitecturerather than the real-timeweight values, the specific weight of each action is a black box, and it is impossible to precisely quantify "how many times more important a bookmark is than a like."
- Starting in November 2025, the Following Feed was also changed to Grok ranking; the algorithm is still continuously evolving, and the conclusions of this document are time-sensitive.
Easily confused adjacent methodologies
- **x-longtail-strategy (long-tail mindset)** is the strategic layer of "accepting the long tail and increasing the probability of reaching the top 10%"; this skill is the mechanistic layer of "how the algorithm scores." The two are complementary: understand the mechanism → adjust your approach → accept the long-tail outcome. Distinguishing signal: ask "mindset/should I give up" → longtail; ask "how does the algorithm work" → this skill.
- **x-cold-start-playbook (cold start)** is the execution layer of "actively heating up in the first few hours to break through the initial pool"; this skill explains "why heating up is effective" (initial engagement signals → algorithm determines quality → pushes into a larger recommendation pool). Distinguishing signal: ask "what should I do" → cold-start; ask "why does the algorithm do this" → this skill.
Related skills
- depends-on
x-cold-start-playbook: The cold-start strategy depends on understanding the algorithm mechanism—first understand how the algorithm scores (OON/candidate isolation/multi-action weighting), then design the six strategies to break through the initial pool. - composes-with
x-longtail-strategy: Algorithm mechanism (why something is recommended) + long-tail mindset (accepting that most posts being ordinary is the norm); after understanding the mechanism, adjust your approach, then use the long-tail mindset to manage expectations.
Audit information
- Verification passed: V1 cross-domain ✓ / V2 predictive power ✓ / V3 uniqueness ✓
- Test pass rate: To be filled (see test-results.md)
- Distillation time: 2026-07-14