I had OpenAI’s Codex CLI read X’s Newly Open Sourced ‘For You’ Codebase. Here is what it found to help you drive engagement.

Published: September 09, 2025

---

# How “For You” Works

• Candidate sourcing: Pulls tweets from multiple sources (in‑network, recs via FRS/UTEG, communities, popular videos, backfills).
• Filtering: Removes/filters spam, unsafe, low‑quality, duplicates, location/visibility misfits, and fatigue.
• Scoring: ML models predict engagement probabilities (per “head”), then a weighted reranker combines them into a final score.
• Reranking: Scores are combined with per‑head normalization, biases/debiasing, and special handling of negative feedback.
• Assembly: Ads and modules inserted; client decorations applied.

# Key Ranking Signals (explicit in code)

• Positive heads (optimize for these):
  • Favorites, replies, retweets.
  • Good clicks (quality click signals), profile clicks.
  • Shares, bookmarks.
  • Dwell on tweet detail.
  • Video: quality views (≥10s when “ten seconds logic” enabled), immersive video views, watch time.
• Negative head (avoid these):
  • Predicted Negative Feedback V2 (mutes, blocks, “show less”, etc.).
• Feature gates/thresholds:
  • Video quality view eligibility often requires video ≥10 seconds.
  • Dwell and VQV can be thresholded or binarized; low dwell can be reduced to zero contribution.
• Weights and normalization:
  • Weights come from feature switches (set server side); model heads are per‑head max‑normalized then linearly combined.
  • Negative weights are filtered by thresholds (constant or normalized) and can still subtract when above threshold.
• Contextual features commonly used:
  • RealGraph/in‑network scores, social proof (liked/retweeted/replied‑by counts + user IDs), author features (verification/creator), media type, embeddings (SimClusters), language match, topic and “Grok” content understanding flags (NSFW, spam, low quality, gore/violence).

# Downranks and Drops (visibilitylib)

• Strong enforcement against: spam, low quality, engagement spamming, copypasta, unsafe/NSFW, crisis/misinformation policies.
• Downrank reasons include spam replies, low‑quality mentions, high spammy content scores, and other health signals.
• Feedback fatigue scorer reduces scores for entities the viewer previously “See fewer” on (authors, likers/followers/ retweeters).

# High‑Leverage Tactics

• Drive “high‑value” engagements
  • Bookmarkable content: deep guides, tools, cheat sheets, templates; explicitly invite saving for later.
  • Shareable posts: concise value + strong hook; include context that travels well.
  • Profile clicks: craft a compelling bio/pinned post; occasionally ask readers to check your profile for related threads/resources.
  • “Good clicks” and dwell: use clear previews that reward opening tweet details (not clickbait); add structured info that takes time to absorb.
• Win video signals
  • Post videos ≥10 seconds to qualify for “video quality viewed”; optimize early seconds to reduce drop‑off.
  • Tight edits, clear audio, subtitles; deliver a payoff before 10–15s to maximize quality watch and watch time.
  • Use immersive-friendly formats where relevant (portrait, clean framing).
• Activate in‑network and social proof
  • Publish when your core audience is online to maximize fast, in‑network engagement.
  • Seed engagement via collaborators/mutuals; early likes/retweets/replies from your network improve candidate viability.
  • Nurture follower quality: interact meaningfully so RealGraph/relationship strength grows.
• Keep language and topic alignment tight
  • Match audience language; avoid mismatched multilingual posts in the same tweet if your audience is mostly one language.
  • Stay within coherent topic clusters; recurring themes help embedding/topic features and candidate recall.
• Avoid negative feedback at all costs
  • Don’t reply‑spam, keyword‑stuff, or post near‑duplicate/copypasta content.
  • Steer clear of misleading/low‑quality bait; health and policy rules can downrank or drop visibility.
  • Watch “See fewer” patterns; if a format triggers fatigue, rotate topic/format.
• Format for consumption and dwell
  • Threads: lead with substance in tweet 1; maintain tight pacing; summarize at end with a save/share prompt.
  • Use clean visuals and readable text over images; avoid heavy image text that seems spammy/low‑quality.
• Timing and cadence
  • Favor windows when your core audience engages; consistency helps both candidate recall and social proof.
  • Don’t flood: fatigue filters and deduplication reduce returns from repetition.

# What Not To Do

• Don’t farm weak clicks; “good clicks” and dwell are what matter, not empty opens.
• Don’t rely on pure “engagement bait” (e.g., low‑effort polls, rote prompts); it risks low quality labels or negative feedback.
• Don’t mass‑reply for exposure; spam‑reply downranks are explicit.
• Don’t post borderline NSFW/violent/gore content; related labels suppress For You ranking.
• Don’t re‑post the same asset repeatedly; dedup + fatigue scorers will discount it.

# Quick Audit Checklist

• Does each post deliver one of: bookmark value, share value, or a profile‑worthy call‑to‑learn‑more?
• For video: ≥10s, strong first 3–5s, captions, clear payoff.
• Thread structure: front‑loaded value, skimmable formatting, explicit takeaways.
• Language/topic: aligned with audience; consistent themes.
• Early engagement plan: alert key mutuals; schedule at high‑engagement times.
• Health: zero spammy patterns; compliant with safety policies.

# Measure And Iterate

• Track rate per impression for: bookmarks, shares, profile clicks, dwell, video quality views/watch time.
• Compare formats: single vs thread; short vs long video; CTAs (save/share/profile).
• Watch audience‑fit signals: who engages early, language distribution, topic resonance.
• Kill formats that produce “See fewer” or high bounce/low dwell.

---

# Candidate Sources

• TweetMixer (ScoredTweetsTweetMixer): General Home recs with reply/directed‑at filters and retweet dedup. Good for broad recall once there’s early engagement.
• UTEG (ScoredTweetsUteg): User‑Tweet Engagement Graph (liked/retweeted by socially close accounts). Strong social proof helps; “reply targeting” can be filtered out (see filters).
• Earlybird In‑Network (ScoredTweetsEarlybirdInNetwork): Recent tweets from followed users; a key starter for early, in‑network engagement.
• Popular Videos (ScoredTweetsPopularVideos): Explore Ranker immersive video recs; favors quality watch/immersion.
• Lists/Communities/Backfill/Cached: Niche or fallback sources that round out supply.

# High‑Impact Filters (what gets removed or discounted)

• Replies/directed‑at misfit
  • QualifiedRepliesFilter: Removes replies by followed authors when the reply targets a user the viewer does not follow and the reply is a retweet, a reply‑to‑reply/retweet, or not a valid reply target. Translation: reply‑chains that look low‑quality or misdirected get filtered.
  • ExtendedDirectedAtFilter: Removes tweets “directed at” someone the viewer doesn’t follow when author is followed by the viewer. Translation: @-shots at non‑followed users from people you follow are reduced.
• Deduplication and hydration
  • RetweetDeduplicationFilter: Multiple RTs of the same original are deduped; original or the highest‑scored retweet wins. Translation: posting original content beats chasing retweet variants.
  • TweetHydrationFilter: Drops tweets that fail Tweetypie hydration. Translation: broken media/metadata kills eligibility.
• Location and fatigue
  • LocationFilter: If your tweet has a location and the user has a location, they must match. Translation: avoid unnecessary/incorrect location tags that shrink reach.
  • FeedbackFatigueFilter and FeedbackFatigueScorer: Viewers’ “See fewer” on your author, likers, followers, or retweeters causes filtering or score discounting for a decay window.
• Content quality/health
  • GrokNsfwFilter, GrokViolentFilter, GrokGoreFilter, GrokSpamFilter: Content understanding labels can drop/ downrank candidates.
• Media hygiene
  • MinVideoDurationFilter/MaxVideoDurationFilter: For video quality signals (e.g., PredictedVideoQualityWatchScoreFeature), ≥10s duration often required when the “ten seconds logic” is enabled.
• Previously‑seen/similar
  • PreviouslySeen* and cluster/media dedup filters: Reduce repeats, near‑dupes, and media repetition.

# Scoring Heads To Optimize (combined in the reranker)

• Positive model heads (maximize via content and structure):
  • Favorites, replies, retweets, “good clicks” (quality click variants), profile clicks, shares, bookmarks, dwell on tweet detail, video quality views (≥10s), immersive video views, watch time.
• Negative head (avoid):
  • PredictedNegativeFeedbackV2ScoreFeature (mutes/blocks/show‑less).
• Reranking logic:
  • Heads weighted via Feature Switch params (ModelWeights.*), per‑head max normalization, and optional bias/debias tensors; negative weights are gated by normalized/constant thresholds and can still subtract when above threshold.

# The “For You Readiness Score” (FYRS)

Use this quick pre‑post checklist to approximate the ranking heads with controllable proxies. Aim for FYRS ≥ 8 to post; do not post if any “Blockers” trigger.

• Value (0–6)
  • +3 Bookmark value: concrete utility that merits saving (guide, template, code, checklist).
  • +2 Share value: concise, portable insight that stands alone if shared.
  • +1 Profile pull: clear reason to view your profile (related thread/pinned resource).
• Dwell/Click Quality (0–5)
  • +2 Strong first line/hook that previews substance (not clickbait).
  • +2 Skimmable structure (short paras, bullets, or thread with summaries).
  • +1 Visual clarity (clean image/video captions, no dense text‑in‑image).
• Social Proof Setup (0–4)
  • +2 Publish in a window where 5–10 core followers are active.
  • +1 Relevance to your recurring topics (consistent cluster/theme).
  • +1 One lightweight nudge to genuine peers (no mass tags/asks).
• Video Only (0–4; if applicable)
  • +2 Duration ≥10s with clear payoff by 5–10s.
  • +1 Captions/subtitles; +1 portrait/immersive‑friendly framing.
• Safety/Health and Fatigue (−8 to 0)
  • −3 Any risk of “See fewer” (you posted same topic ≥3x this week with weak engagement).
  • −3 Looks like reply/directed‑at bait (reply to non‑followed or @ aimed at non‑followed without context).
  • −2 Near‑duplicate to recent post or recycled media.
• Integrity Blockers (hard stop)
  • NSFW/violent/gore/spammy flags likely; misleading claims; broken media; unnecessary location tag; fails hydration in preview. If any true, do not post.

Target: FYRS ≥ 8 and no Blockers.

# How To Score Yourself (fast pass)

• If thread: lead tweet delivers standalone value; later tweets deepen value; end with summary + “save/share”.
• If single image: conveys an idea that invites a save or a share; caption adds context.
• If video: ≥10s, substantive takeaway by 5–10s, captions on, cover frame communicates the value.

# Source‑Fit Micro‑Checks

• In‑network kickstart (Earlybird/In‑Network): Will at least 5 core followers likely engage in 15–30 minutes?
• UTEG/social proof: Is the topic aligned to people who recently liked/retweeted you? Avoid low‑quality reply chains or @‑shots to non‑followed accounts (tripped by QualifiedRepliesFilter/ExtendedDirectedAtFilter).
• Popular video lane: If video, does it pass the 10s quality view bar and feel immersive?

# Operational Guardrails

• Post cadence: Avoid posting the same asset/angle repeatedly (fatigue + dedup filters).
• Location: Don’t attach location unless it’s critical and broadly relevant to your audience.
• Retweets vs original: Prefer original content; retweets of the same origin are deduped.
• Language/topic: Keep consistent language and coherent topics; avoid mixed‑language posts that don’t match audience.

# How To Make It Testable

• Track per‑post “high‑value” rates you can infer:
  • Bookmarks, shares, profile clicks, dwell proxy (detail expands), video 10s+ views/watch time.
• After 5–10 posts, calibrate FYRS thresholds:
  • If posts with FYRS ≥ 9 outperform median impressions per follower and show higher bookmark/share rates, your rubric is aligned.
  • Tighten negatives if “See fewer” or muted patterns rise.