---

name: x-for-you-growth description: Grow on X (Twitter) using the open-source For You / Phoenix ranking signals. Use when the user wants more reach, impressions, followers, or replies; asks how the X algorithm works for creators; needs draft audits, hooks, threads, reply strategy, posting cadence, or a "why did my reach die" diagnosis grounded in xai-org/x-algorithm weights—not engagement pods or spam tactics.

X For You Growth

Turn the open-source X For You algorithm (xai-org/x-algorithm) into practical drafting, posting, and engagement coaching. Prefer primary sources (repo README + home-mixer/params/param.rs) over third-party blogs. Production defaults are mirrored into the repo on a cadence; re-check weights before giving numeric advice.

This skill does not post, spam, buy engagement, or game safety systems. It drafts and coaches. Confirm before any live publish.

When to Use

• User wants more reach, impressions, followers, or replies on X/Twitter
• User asks how the X / For You / Phoenix algorithm works for creators
• Drafting hooks, threads, reply strategy, or a posting cadence for X
• Auditing a draft or account against ranking signals
• Comparing formats (text, image, video, quote, thread, link)

Don't use for: LinkedIn-only growth, ad buying playbooks, engagement pods/bots, mass follow-unfollow, or circumventing visibility/safety labels.

Prerequisites

• Optional: fetch latest weights from the public repo before numeric claims
  • https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
  • https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
  • Scoring math: home-mixer/scorers/ranking_scorer.rs
• See also references/param-baselines-2026-08.md in this skill (snapshot; may go stale)

Core model (what actually ranks)

Sources (parallel)
  Thunder          → in-network (accounts viewer follows), recent posts
  Phoenix retrieval→ out-of-network two-tower similarity
  SimClusters      → community / engagement-cluster OON
       ↓
Pre-filters (age ≤ ~48h, blocks/mutes, muted keywords, already seen/served, duplicates, OON retweet/reply rules, …)
       ↓
Phoenix ranker     → P(favorite), P(reply), P(repost), P(quote), P(share…), P(dwell…), P(follow), P(negatives…)
       ↓
Final Score ≈ Σ (weight_i × P(action_i))
       ↓
Adjustments: author diversity decay · OON discount · new-author/cold-start boost · VM-ranker diversity reorder
       ↓
Visibility filtering can DROP / interstitial after rank (spam, safety, labels)

Implications:

1. You need candidate selection (followers + OON retrieval/clusters) and a high weighted score.
2. The model optimizes predicted viewer actions, not abstract "quality."
3. Negative feedback (report/mute/not interested/block) crushes score.
4. Content older than the age filter (~48 hours) is out of For You consideration.
5. Multiple posts from the same author in one feed are decayed (diversity).
6. Out-of-network posts are discounted vs in-network (need stronger predicted engagement to surface).

Published weight baselines (mirror ~2026-08-12)

Re-fetch param.rs before treating these as current. Relative to favorite = 0.5:

Signal	Default weight	≈ × favorite	Creator takeaway
Share via copy link	20.0	40×	Make something people save/forward
Bidirectional-follow reply boost	+15.0 on reply path	huge	Mutuals replying (and deep reply chains with people who follow each other) are elite
Reply	5.0	10×	Optimize for conversation, not likes
Quote	5.0	10×	Quotable, disagreeable, specific takes
Share via DM	5.0	10×	Private "you need to see this" value
Follow author	4.0	8×	Profile-worthy posts convert lurkers
Share (general)	2.0	4×	Shareable one-liners / frames
Retweet	1.0	2×	Amplification still matters
Favorite (like)	0.5	1×	Weak alone; vanity metric
Click	0.4	0.8×	Curiosity / expand still helps
Open link	0.2	0.4×	External links are weak positives in-score (and often hurt stay-on-platform behavior in practice)
Photo expand / video open / VQV	~0.05	tiny	Media helps via dwell/expand, not as a magic multiplier alone
Continuous dwell time	0.004	tiny per unit	Long readable content still compounds via attention heads
Not dwelled	-0.02	small neg	Instant skips hurt
Not interested	-43.2	catastrophic	Topic mismatch / bait fatigue
Block author	-31.2	catastrophic	Polarizing-to-hostile is risky
Mute author	-58.8	catastrophic
Report	-234.0	nuke	Spam, scams, harassment, policy breaks

Also important (not a simple weight):

• Author diversity: decay=0.5, floor=0.25 → 2nd post from same author heavily attenuated in one feed build.
• OON factor: 0.75 (topic OON 0.5) → discovery posts need stronger predicted engagement than follower feed posts.
• Cold-start / new-author boost exists for low-impression authors (quality + early velocity still required).
• OON-style discount can also hit in-network replies/reposts when that rescoring flag is on.

Third-party "author reply = 150 likes" lore comes from older hand-tuned systems. Current public Phoenix weights emphasize reply/quote/share/copy-link/follow and mutual-reply boosts — treat old absolute multipliers as directional only unless you re-derive them from live param.rs + scorer code.

Procedure

1) Refresh algorithm facts when stakes are high

If the user needs numeric claims or the repo may have moved, re-read param.rs and the README scoring section.

Completion: you can state last-synced defaults and whether reply/share/report weights changed.

2) Diagnose the growth goal

Capture: niche, audience, current follower band, posting frequency, best/worst recent posts, link-in-bio needs, Premium or not, primary KPI (followers vs site clicks vs authority).

Completion: one-sentence goal + constraint list.

3) Account hygiene (candidate + trust layer)

Coach, do not automate abuse:

• Clear niche bio + pinned proof post (raises follow conversion when P(follow) fires)
• Healthy follow graph (engage real peers; avoid mass follow/unfollow patterns that look inauthentic — behavioral detection / cred systems exist)
• Check Under the Hood labels if reach suddenly dies: https://x.com/i/under_the_hood
• Remove spammy media, engagement bait templates, misleading claims that drive reports/mutes

Completion: short account checklist with pass/fail.

4) Draft for high-weight actions

When writing or reviewing a post, force these properties:

1. Reply bait without empty "thoughts?" — specific claim, prediction, tradeoff, or sharp question a niche peer can answer in one line.
2. Quotable core — one sentence others can quote-tweet with their own spin.
3. Copy-link / DM worthiness — "send this to X role" utility (checklist, non-obvious number, crisp framework).
4. Follow reason — reader should know what they get by following (series, beat, POV).
5. Dwell — short paragraphs, concrete nouns, media that rewards expand; threads when depth beats a blob.
6. Minimize negatives — no scrape-bait, fake urgency, tag-spam, or topic bait-and-switch.

Format defaults:

Format	Use when	Notes
Single text	Hot take / news reaction	Highest reply rate if opinionated and specific
Image + text	Diagram, screenshot, before/after	Photo expand is a real head; keep text self-contained
Native video	Demo, talking head, screen record	Optimize completion / VQV; native upload, not YouTube link as the post
Thread (3–7)	Teaching / story	First tweet is the ad; each line earns the next; end with a reply prompt
Quote-tweet	Commentary on a primary source	Your frame must stand alone if parent is collapsed
Link	Distribution of off-platform asset	Prefer value in main post, link in first reply (or later) so ranking is not wedded to weak open_link

Link policy: score weight for open-link is low; many creators still see weaker distribution with external URLs in the root post. Default: native insight first, URL in reply unless user insists otherwise.

Completion: draft has hook, body, CTA-to-reply or follow, and optional media note.

5) Timing and first-hour ops (velocity)

For You still needs early positive actions so OON retrieval/social proof can fire:

1. Post when core followers are online (use their history / analytics if available).
2. Block 20–40 minutes after posting for real replies (not emoji spam).
3. Reply in ways that add information (extends dwell + more reply probability on the thread).
4. Soft-seed: reply thoughtfully on 3–5 niche posts before/after your post (human, on-topic — not "great post!" spam).
5. Do not blast the same link 10 times; author diversity + spam labels will punish.

Completion: post plan includes time window + reply SOP.

6) Cadence vs diversity penalty

• Prefer fewer stronger posts over many near-duplicates the same hour.
• There is author diversity decay inside a single feed assembly: back-to-back near-identical posts cannibalize each other.
• There is a ~48h age cutoff for For You candidates — evergreen recycling only helps if reposted/refreshed as a new candidate with new engagement, not as infinite old inventory.

Completion: weekly calendar with spacing, not a firehose.

7) Growth loops that match the math

Priority order for organic growth:

1. Conversation density on your posts (replies + your substantive replies)
2. Quotes and copy-link shares (high weights)
3. Follow conversions from profile-worthy posts
4. In-network amplification from mutuals (bidirectional reply boost path)
5. Cluster belonging via consistent niche (SimClusters + Phoenix history)
6. Likes last (lowest positive engagement weight)

Avoid: pods, reply-guy spam under huge accounts, engagement bait that triggers not-interested/report.

8) Deliverables by request type

User ask	Deliver
Explain algorithm	Short pipeline + weight table + caveats + links to repo
Write a post	2–3 variants optimized for reply/quote/share; hook graded
Content audit	Score draft on reply, quote, share, follow, dwell, negative risk
30-day plan	Niche pillars, cadence, reply hours, metrics to watch
Thread	Outline + tweet-by-tweet with completion CTA
"Why did reach die?"	Checklist: age, labels/UTH, spam patterns, link-only posts, diversity spam, topic drift, negatives

Quick drafting checklist

• Specific audience + one clear POV in line 1
• Invites a real reply (not generic engagement bait)
• One quotable line
• Useful enough to copy link / DM
• Media native if used; text works if media fails
• External link deferred to reply (default)
• No multi-post spam in the same window
• Plan to answer replies in first hour
• Topic matches account niche (reduces not-interested)
• Nothing that invites mass report/mute

Metrics that matter (coach the user)

Watch for 24–48h per post:

1. Replies and reply rate (primary)
2. Quotes + shares (incl. soft signals if available)
3. Profile visits → follows
4. Bookmark / share behavior if shown in analytics
5. Dwell proxies: detail expands, video completion
6. Negative: "not interested," blocks, reports (if visible), sudden distribution cliffs → Under the Hood

Deprioritize raw like count as the success metric.

Pitfalls

1. Stale third-party weight tables. Always prefer live param.rs. Blogs often mix 2023 hand-tuned weights with 2026 Phoenix.
2. Like farming. Favorite weight is low; reply/quote/share/follow dominate.
3. Link-in-root habit. Weak score contribution; often worse distribution in practice.
4. Thread-as-essay dump. First tweet must earn the click; each hop needs a reason.
5. Engagement bait ("like if you agree") → not-interested and mute risk.
6. Spam replies under celebrities. Short-term impressions, long-term cred/label risk.
7. Ignoring negatives. One report weight dwarfs many likes.
8. Posting 8 near-duplicates daily. Author diversity + filters + fatigue.
9. Topic whiplash. Phoenix user history + SimClusters reward coherent niches; random viral formats train the wrong retrieval neighbors.
10. Promising guaranteed virality. System is personalized per viewer; same post scores differently for every user.

Verification

• Algorithm summary cites repo paths and notes param sync uncertainty.
• Every growth recommendation maps to a named signal (reply, quote, share_via_copy_link, follow_author, OON discount, age filter, diversity decay, VF/labels).
• Drafts pass the Quick drafting checklist.
• If user reports dead reach: include Under the Hood + hygiene checklist before "just post more."

Sample usage

Worked prompts and abbreviated ideal answers live in references/sample-usage.md.

Full freeze-dried agent outputs from a live skill test: references/sample-output-live.md.

Prompt	What you should get
"How does For You actually rank posts?"	Pipeline + weight hierarchy + 48h/OON/diversity caveats + repo links
"Rewrite: Excited to share my blog… link in bio"	2–3 variants optimized for reply/quote/share; URL moved to reply; signal map
"Audit this draft before I send"	Scored rubric (reply, quote, copy-link, follow, dwell, negative risk) + fix list
"First-hour plan after I post at 8am"	Timed SOP (seed replies, stay present, no spam duplicates)
"Impressions collapsed — same niche"	Ordered diagnosis (Under the Hood → bait/links → topic drift → cadence)
"7-day plan for a small personal-AI account"	Day-by-day format + target signal + KPI focus on replies/follows

Quick invoke lines

• Explain X For You ranking like a creator coach.
• Rewrite this tweet for replies and quotes; put the link in a reply.
• Audit this draft against Phoenix weights.
• My reach died — run the diagnosis checklist.

Reference links

• Repo: https://github.com/xai-org/x-algorithm
• README scoring section (How It Works → Scoring and Ranking)
• Params: home-mixer/params/param.rs
• Scorer: home-mixer/scorers/ranking_scorer.rs
• DeepWiki index: https://deepwiki.com/xai-org/x-algorithm
• Under the Hood: https://x.com/i/under_the_hood
• Local snapshot: references/param-baselines-2026-08.md
• Sample prompts/outputs: references/sample-usage.md
• Live freeze-dried run: references/sample-output-live.md