X didn’t open source the algorithm for fun.
They published the system that decides what gets shown in the For You feed and what gets filtered out over time. A Grok-based model that predicts how people will react to posts and ranks them accordingly.
If you’re a creator, this is leverage. Understand what the model optimizes for and you stop guessing. You start posting with intent.
1. The mental model (no math, just structure)
From the repo, the For You feed works like this:
Thunder → recent posts from people you follow.
Phoenix Retrieval → posts from people you don’t follow but are embedding-similar to your interests.
Phoenix Scorer (Grok-based transformer) → evaluates candidates using:
Your engagement history (likes, replies, clicks, watches)
The candidate post itself
It predicts actions:
P(like), P(reply), P(repost), P(quote), P(click), P(profile_click), P(dwell), P(follow_author)
And negative actions: P(not_interested), P(mute_author), P(block_author), P(report)
Weighted Scorer → combines those probabilities into a single score using weighted sums.
Positive actions add score. Negative actions subtract score.
Author diversity + filters → limits repeated authors and removes blocked, muted, NSFW, or ineligible content.
The system returns the top K posts.
Two important details:
Hand-engineered features are removed. The model learns directly from behavior.
Candidates do not attend to each other. Each post is scored independently against the user.
So the goal as a creator is simple.
Be high upside. Be low risk. Repeatedly.
2. The five levers you actually control
Lever 1: Optimize for multi-action engagement, not likes
The model does not care about likes in isolation. It predicts multiple actions and weights them.
Replies matter more than likes.
Reposts and quotes matter for distribution.
Profile clicks and follows matter for identity.
Dwell matters because it signals attention.
Media expands and video views matter because they show intent.
Action items:
Write posts that invite replies, not agreement.
Ask questions that require thinking, not yes/no answers.
Include one concrete takeaway per post so it can be saved or shared.
Think bookmarkable plus discussable.
Not agreeable and forgettable.
Lever 2: Treat negative feedback as toxic
The model explicitly predicts negative actions.
Those actions directly reduce score.
Ragebait may spike impressions short term.
It increases mute, block, and not interested actions long term.
That damages future distribution to similar users.
Action items:
Avoid repetitive clickbait with no payoff.
Avoid bait-and-switch threads.
Avoid provoking audiences that are not your target.
Before posting something provocative, ask:
Is this on-topic or am I bored?
If people start saying “why do I keep seeing this,” that is a warning signal.
Adjust.
Landmine:
An audience that dislikes you still trains the model.
Just against you.
Lever 3: Respect author diversity (you compete with yourself)
The system penalizes showing too many posts from the same author in one feed.
If you post frequently and most posts underperform:
You waste impressions.
You crowd out your own best posts.
Action items:
Post fewer, higher-intent main posts.
Aim for 2 to 4 strong posts per day.
Use replies for volume. Replies do not hit author diversity the same way.
Space your pillar posts so the system can test each one properly.
Lesson:
Volume helps. Unfocused volume hurts.
Lever 4: Be discoverable in Phoenix Retrieval (out of network)
Out-of-network reach is how accounts grow.
Phoenix Retrieval uses a two-tower embedding model:
One tower encodes users.
One tower encodes posts.
Similar embeddings get matched.
This means:
Topic consistency matters.
Audience quality matters.
Action items:
Pick a lane.
Make at least 70 percent of your content about the same topic cluster.
Focus on getting engagement from people already deep in that niche.
Avoid scattering topics.
DeFi today, politics tomorrow, sports after that.
That makes you a weak candidate everywhere.
Lesson:
The system uses who engages with you to decide who should see you next.
Curate your topic and your audience.
Lever 5: Train your audience graph intentionally
The model consumes engagement as a sequence.
What users do matters more than what they say.
For readers:
Your feed reflects your behavior.
For creators:
Your future reach is shaped by who engages with you now.
If most engagement comes from:
Airdrop farmers
Engagement pods
Low-signal replies
The system learns to show you to more of them.
Action items:
Block or mute obvious bot clusters.
Stop incentivizing low-effort engagement.
Write explicitly for the audience you want more of.
Example framing:
“This is for people actually working on X.”
“If you are serious about Y, read this.”
Reply to the right accounts. Not just any account.
Landmine:
Low-signal engagement produces low-signal reach.
3. Concrete playbook
Daily
One core post:
Dense with value for your niche.
Includes a natural reply hook.
Worth saving or referencing.
One bridge post:
More accessible.
Still on-topic.
Designed to attract new but relevant users.
Replies:
50 to 100 meaningful replies.
Focus on threads where you add clarity.
Each reply is mini-content and a training signal.
Audit:
Watch for drops in replies or reposts.
Watch for “why am I seeing this” comments.
Adjust quickly.
Weekly
Run one format experiment:
Carousel
Short video
Compact essay
Track:
Dwell
Profile clicks
Follows per impression if possible
Clean your audience:
Remove spam replies.
Mute or block farming rings.
Check topic alignment:
Scroll your timeline.
Ask if your account is easy to categorize.
You do not need to be a robot.
But your dominant signal must be clear.
4. Landmines to avoid
Ragebait:
Short-term reach. Long-term suppression.
Engagement pods:
Repetitive low-quality engagement trains the wrong audience.
Off-topic viral posts:
Traffic does not stick and weakens embeddings.
Flood posting mid content:
Trains the model that your posts are ignorable.
Template slop:
Low dwell, low replies, low reshares equals background noise.
5. The core lesson
Most people talk about beating the algorithm.
That framing is wrong.
The system is pattern matching on behavior.
Your behavior and your audience’s behavior.
You are constantly teaching it:
What you post
Who engages
How they engage
What you engage with
You are not fighting the system.
You are shaping how it understands you.
Once you internalize that, the For You feed stops feeling random.
It becomes predictable.
Not through hacks.
Through consistency.
If you are a creator in 2026, this is the job:
Curate your topics
Curate your audience
Curate your behavior
Do that and the system does exactly what it is designed to do.
Show your content to people likely to care.
Your goal is simple.
Be the account the model would regret ignoring.
