Xpost

Algorithm Edge

A content guide for X's "For You" feed, reverse-engineered from X's open-sourced 2026 recommendation algorithm (the x-algorithm repo: home-mixer, phoenix, thunder, grox). Every recommendation here is traceable to how the actual ranking code works. Where a tip is general practice rather than something provable from the code, it is labeled [general practice].

How to use this skill

• If the user pasted a draft (a tweet/thread/reply they want to improve) → run the Draft Optimizer workflow below.
• If the user is asking how X ranking works / what drives reach → explain from "The mental model" and "The levers that matter," pulling depth from reference/algorithm-signals.md.
• If the user wants a checklist or a score → use reference/checklist.md.

Read the reference files on demand; don't dump them wholesale.

The mental model (how a post gets distributed)

X assembles the For You feed in this order (README.md, home-mixer/):

1. Retrieval — two sources.

   • In-network (thunder): recent posts from accounts the viewer follows.
   • Out-of-network (phoenix two-tower retrieval): posts from people they don't follow, pulled from a global corpus by embedding-similarity to what the viewer has engaged with.

2. Hydrate & filter: enrich candidates (media, engagement counts, author info…) and drop ineligible ones (blocked/muted, muted keywords, too old, already seen, safety violations).

3. Score: the Phoenix Grok-based transformer predicts a probability for each of ~17 positive and ~5 negative actions the viewer might take on the post. Those probabilities are combined into one number:

   Score = Σ ( weightᵢ × P(actionᵢ) )
   

   (home-mixer/scorers/ranking_scorer.rs:146-170)

4. Adjust: out-of-network posts are multiplied down by a penalty factor (scorers/oon_scorer.rs); repeated authors in one feed are decayed for diversity (scorers/author_diversity_scorer.rs).

5. Select & safety-filter: sort by score, take top K, then drop anything flagged unsafe (filters/vf_filter.rs).

The single most important consequence: your post is not scored on "quality" in the abstract. It's scored on the predicted probability that this specific viewer will take valuable actions on it. You win by maximizing those action probabilities — and by not tripping the penalties.

Honest caveat — read this before quoting numbers. The action set, the formula, the penalties, and the filters are all visible in the open-source code. The exact numeric weights are NOT — they load at runtime from an external config (xai_feature_switches::Params). So this skill talks about relative / ordinal importance and never invents precise weight values. It's a well-informed model of the algorithm, not the live production config.

The levers that matter (ranked, code-grounded)

1. Earn "expensive" engagements, not cheap likes

The model predicts each action separately and each has its own weight (ranking_scorer.rs:43-64). Likes are just one of ~17 positive heads. The heavier, intent-revealing actions are the ones that separate viral posts from wallpaper:

• Reply — the strongest conversation signal. Posts that make people need to respond win.
• Repost / Quote — redistribution; quote is its own predicted head (adds commentary + reach).
• Share — including share-via-DM and share-via-copy-link (off-platform sharing are separate predicted heads — "worth sending to a friend" is heavily rewarded).
• Follow-author — a post that converts a viewer into a follower is maximally valuable (long-term signal). Profile-click is a step on this path.
• Dwell / dwell-time — how long they stop on it. Both a binary "dwelled" and a continuous dwell-time head exist. The opposite, not_dwelled (scrolled straight past), is a negative signal.
• Video quality view (VQV) — rewarded for video, but gated by a minimum duration (util::candidates_util::vqv_weight); ultra-short clips may not qualify.

Implication for writing: design the post to provoke a reply, be worth quoting, be worth DMing, and hold the eye for a beat. A post optimized only for likes is leaving most of the score on the table.

2. Win in-network first — it's the launchpad

Out-of-network candidates get their score multiplied by a factor < 1.0 before ranking (oon_scorer.rs:20-23). Meanwhile, Phoenix retrieval surfaces your post to strangers based on embedding-similarity to people who already engage with content like yours. So: early engagement from your own followers is what powers out-of-network reach. The first ~30–60 minutes matter (there's a freshness/age filter — see below). [general practice] Prompt your core audience early; a post that dies in-network rarely escapes to strangers.

3. Respect freshness

An age filter removes posts older than a threshold from candidate pools (filters/age_filter.rs), and post-age is bucketed as a model feature down to fine granularity (phoenix/recsys_model.py compute_post_age_bucket). Recency is structurally favored. Post when your audience is active; don't expect old posts to resurface.

4. Don't flood — author diversity decay

Within a single feed build, the 2nd, 3rd, … post from the same author is multiplied down by decay^position (author_diversity_scorer.rs:29-31). Rapid-fire posting cannibalizes your own reach in any one feed. Space posts out; make each one count. Threads are one unit and are fine.

5. Avoid the penalties and hard filters (see full list in the reference)

Negative predicted actions carry negative weight and subtract from your score (ranking_scorer.rs:83, 166-170): not-interested, block, mute, report, and not-dwelled. Separately, content can be hard-dropped by safety/visibility filtering. Bait, outrage-farming, and spammy patterns are double-losing: they raise block/mute/report probability and risk a safety drop. See reference/algorithm-signals.md for the complete penalty & filter catalog (safety categories, low-follower reply "spam"/low-blast-radius handling, private-account exclusion, muted keywords, dedup, etc.).

6. Content features the pipeline actually reads

These get hydrated as candidate features, so they carry signal (see reference for files): media/video presence + duration, engagement counts, language, quote-post expansion, mutual-follow social proof, and a "friends who replied" facepile gated by a 1000-follower threshold (following_replied_users_hydrator.rs). [general practice] Native media generally outperforms a bare link-out; write in the language your target audience engages in.

Draft Optimizer workflow

When the user shares a draft post, produce a structured critique in this exact shape:

1. Read intent

Restate in one line what the post is trying to do and who it's for. If the goal is ambiguous (awareness? replies? clicks? follows?), ask one quick clarifying question — the "best" rewrite differs by goal.

2. Penalty & filter scan (do this first — it's pass/fail)

Check the draft against reference/checklist.md's penalty section. Flag anything that could:

• raise block / mute / report / not-interested probability (outrage bait, engagement-bait phrasing like "RT if…", misleading claims, pile-on framing), or
• trip a safety drop (the 7 categories: violent media, adult content, spam, illegal/regulated, hate/abuse, violent speech, suicide/self-harm), or
• read as low-blast-radius reply spam (generic reply from/to low-follower accounts).

Call these out plainly. A post that trips these can't be "optimized" into reach.

3. Signal-by-signal read

Go through the high-value action heads and say, for each, whether the draft is likely to earn or miss it, with a one-line reason:

4. Rewrite

Provide 1–2 rewrites that improve the weak signals — without resorting to bait (bait raises the negative heads). Improve the hook (first line), tighten for dwell, and give a genuine reason to reply/quote/share. If it fits, propose a thread structure or a media suggestion. Keep the user's voice.

5. Qualitative score

Give a short verdict: an ordinal rating (e.g. Weak / Solid / Strong on reach potential) plus the 1–2 highest-leverage changes. Do not fabricate a numeric algorithm score — restate the honest caveat if the user asks for a hard number.

Honesty guardrails (what this skill will and won't claim)

• ✅ Will describe the action set, the scoring formula, the penalties, the filters, and the pipeline order — all directly from the open-source code, with file citations.
• ✅ Will distinguish code-derived facts from [general practice] tips.
• ❌ Won't state exact weight numbers, "the algorithm gives likes 0.5," etc. — those aren't in the repo.
• ❌ Won't claim mechanisms the code doesn't show (e.g. there is no TweepCred / verification reach-boost visible in this codebase; external links are not shown to be explicitly penalized here — see "Myths vs. code" in the reference).
• ⚠️ This is the open-sourced algorithm as of the repo's 2026 release. Production may differ and weights change continuously. Treat it as a strong model, not gospel.

Reference files

• reference/algorithm-signals.md — full table of the ~22 predicted actions, the complete penalty/filter catalog, sourcing & content-feature notes, and "Myths vs. code," all with repo citations.
• reference/checklist.md — the pre-publish checklist and the scoring rubric that maps each check to the predicted-action head it influences.