X Engagement

X Engagement Playbook

Expert at building authority and engagement on X (Twitter) through distribution engineering, algorithm-aware content design, and conversation tactics — grounded in the xai-org/x-algorithm codebase at snapshot bc8e5f0 (2026-08-28), which raised reply-spam/ranking eligibility to 120k followers, re-enabled a small binary-dwell weight, and simplified VMRanker to DPP-only on top of the 2026-08-21 80k / SID-slate and 2026-08-14 scoring-semantics releases.

Overview

• Two systems, not one — ranking decides your order; visibility filtering decides whether you're eligible at all. Most reach problems are the second one.
• Published weights — the blend weights are public as of 2026-08-13; the skill cites them with file:line rather than inferring a hierarchy
• Weights multiply P(action), not counts — upstream (2026-08-14) explicitly rejects "1 report cancels N likes." Report is rare (>1000× less base rate than like), so its weight is large so the prediction can move the score. Mass-report brigades are personalized and mostly affect similar users; only Home-Timeline-served actions count
• Account standing gates everything — spam/slop labels drop every post from out-of-network for 30 days while followers see you normally. A new panda_reports_embedding_v10_rough_spam rule can non-permanently suspend for PlatformManipulation
• Mutuals compound — a mutually-followed author gets +15.0 on reply weight for root posts; follow-graph bool features are wired into Phoenix candidate/history tensors
• Conversation leverage — replies are Grok-scored 0–3; spam/reply-ranking eligibility now covers threads where target+root are ≤120k followers (was 80k; was 30k; was 15k). The 60s scoring rate-limit is gone, and a worse ranking score overwrites a better one
• Dwell is small, not zero — DwellWeight is 0.05 (was 0.0); Phoenix scoring aggregation is DENSE_WITH_LONG_DWELL. Still far below reply/quote (5.0). vqv is now 0.0
• VMRanker is DPP-only — ranking_scorer no longer computes SID fields into the reranker request. SlateContext still carries SID + new Phoenix recon_* reconstruction-similarity features from the proto
• Following blocks quotes/RTs — Following timeline hydrates blocked-by on quoted and retweeted authors and drops those candidates
• Monetization alignment — Original Content Rewards pays on verified impressions on original posts in the Home Timeline
• Freshness hardens after 14d — stale posts zero engagement-count features in Phoenix when the flag is on; old viral posts lose raw-count advantage
• Cold start is Home-Timeline impressions — view_count_on_home replaces raw view_count; Thompson-sampling TopK (still off by default) is 2, not 5

Core Principles

1. Don't Get Dropped Before You Get Ranked

Visibility filtering runs a set of rules that fire only for out-of-network recommendations — spam-high-recall, do-not-amplify, abusive, NSFW, compromised. Many are account-level. If OON reach dies while follower engagement holds, that's the cause, not your hooks. Jurisdiction filters (e.g. Brazil2026ElectionFilter, list updated 2026-08-27) can also remove listed authors from For You unless the viewer follows them.

2. Negative Signals Matter — But Read the Math

report −234.0, mute_author −58.8, not_interested −43.2 versus a top positive of 20.0. Those numbers weight predicted probabilities for this viewer, not raw engagement tallies. Do not say "one report cancels ~47 replies." Still: content that invites mute/block/report is expensive, and agatha/ scores blocks/reports relative to favorites as a durable account label.

3. Hook → Hold → Earn a Real Action

dwell is weighted 0.05 (same tier as photo expand) and cont_dwell_time 0.004; not_dwelled is −0.02. Attention is still the precondition, not the payoff. Hold the reader to earn a reply, quote, DM share or follow — those are what score. Do not treat 0.05 as a reason to write long posts for dwell's sake.

4. Follows Compound, Even Though They Aren't Top-Weighted

follow_author is 4.0 — below reply, quote and DM share. But it's the only action that changes every future impression, moves you in-network past the 0.75 OON discount, and feeds the PageRank credibility that buys enforcement leniency. Design for it for those reasons.

5. Post First, Drive Traffic Second

Always post your own content to your profile first, then reply into relevant threads to pull traffic back. Never let your best content die inside someone else's thread — reply impressions don't monetize either.

6. Reply Quality Over Volume — This Is The Riskiest Lever

Replies are Grok-scored 0–3, and the scorer now sees follower counts. Below ~1,000 followers spam scrutiny is elevated, and fast_reply_spam_post carries a 30-day SpamHighRecall label. bdsm/ reads posting cadence directly. Reply-spam / reply-ranking tasks now cover threads up to ≤120k followers on target and root. Every reply can be scored immediately (no 60s dedupe), and a later worse score replaces a better one. Five excellent replies beat fifty mediocre ones by a wide margin.

7. Volume and Repetition Both Decay

Author diversity: your 2nd post in a feed load keeps 62.5%, your 3rd 43.75%. VMRanker separately demotes posts similar to their neighbours via a determinantal point process (DPP). Phoenix reconstruction-similarity (recon_cos_milli, recon_count_above, recon_gap_above) is now an explicit slate-context feature. Post less, and don't rephrase yourself or flood one topic cluster.

8. Originality Is Attributed, Paid, and Enforced

Original Content Rewards pays on qualified impressions: Premium viewers, Home Timeline, original posts. Separately, llm_slop_user and llm_slop_post are real enforcement labels with 30-day TTLs. Generic AI-shaped output is a named, penalized category.

9. Ship While Fresh

Phoenix can mark posts older than ~14 days as stale and zero their engagement-count features (IS_STALE_POST14D). Design for timely distribution; do not rely on ancient high-count posts to keep ranking on raw fav/view features alone. Cold-start explore uses Home Timeline impression counts (view_count_on_home), not global views.

10. Account-level NSFW bleeds into ranking features

Phoenix receives an author-NSFW safety bit (SAFETY_BIT_AUTHOR_NSFW via nsfw_author_phoenix). Adult-marked accounts are not only a visibility problem — they are a model-input problem on every non-retweet candidate.

Documentation

• Scoring Weights - The published blend weights with file:line citations, correct P(action) semantics, bidirectional-follow boost, OON factors, author diversity, dwell 0.05 / vqv 0.0, params off by default
• Algorithm Signals - Which signals exist, candidate sources, network alignment, facepile, stale-post / cold-start notes, how to prioritize
• Visibility Filtering - ALLOW/INTERSTITIAL/DROP, the out-of-network-only drop rules, Brazil 2026 election filter, muted-keyword Following expansion, Following blocked-by quote/RT drop
• Account Standing - agatha, user-cred-v2 PageRank, bdsm behaviour model, the enforcement label chain and its 30-day TTLs
• Content Quality Screening - Banger Screen outputs, reply spam buckets, 0–3 reply rubric, the ten safety categories
• Content Strategy - Hooks, clusters, attention, diversity decay, freshness
• Conversation Tactics - Reply scoring, spam risk (≤120k), thread hijacking, social proof
• Authority Building - Follow triggers, share signals, network alignment, positioning
• Monetization - Original Content Rewards: eligibility, qualified impressions, originality rules, payout mechanics (sourced from X help pages, not the codebase)
• Content Ideas - High-performing templates
• x-algorithm README (upstream) - Cached upstream README
• upstream/ - Verbatim cached source files the analysis above is derived from

Quick Workflow

1. Check standing first if reach has dropped — Under the Hood shows the labels actually on your account; don't theorize about hooks until that's clean
2. Draft content with a strong hook (controversial truth or myth-busting)
3. Ask: could this make someone mute, block or report? Negatives still dominate risk, even after correcting the weight-ratio myth
4. Ask: does this earn a reply, quote, DM share, or follow? Those are what score
5. Check it reads as human — not templated, not your last post rephrased, not the same cluster as the post you just shipped
6. Post to your profile first
7. Wait 10–30 minutes, then find active threads (20–200 likes, your topic)
8. Reply with quality, at human pace — extend the idea, don't self-promote; quality still matters on mid-tier threads up through ~120k

Content Formula

clean account standing (no OON drop labels)
+ topically legible (VLM can categorize it)
+ nothing that invites mute / block / report
+ a reason to reply, quote, or DM it to one specific person
+ a reason to follow (unique insight + consistent content identity)
+ original and human-sounding (not slop-labelled)
+ timely enough that engagement features still count
+ not a near-duplicate of your last post or the same cluster
= algorithmic reach → qualified impressions → payout

Anti-Patterns

• High-volume or fast-cadence replies (fast_reply_spam_post, 30-day SpamHighRecall)
• Generic / templated / AI-shaped content (llm_slop_user, llm_slop_post)
• Rephrasing your own last post (VMRanker DPP demotion, SpamEmbeddingMajorityPoster)
• Flooding one topic cluster in a short window (Phoenix recon_* similarity on slate context)
• Burst posting on a mechanical schedule (bdsm/ reads inter-action timing)
• Deliberate antagonism (agatha scores blocks/reports relative to favorites)
• Posting too frequently (author diversity decay attenuates each successive post)
• Dropping naked links (MALICIOUS_URL_DROP keys on the link, and it can fire retroactively)
• Risky avatar / banner imagery (NSFW_AVATAR_IMAGE_USER_DROP costs you all OON reach)
• Adult-labelled account state leaking into Phoenix via nsfw_author_phoenix
• Posting before building your own content
• Leading with theory instead of concrete examples
• Building a reply-heavy strategy for income (reply impressions don't monetize)
• Aggregating others' content with thin additions (devalued under Original Content Rewards)
• Treating dwell 0.05 as the payoff (reply/quote still 5.0)
• Chasing vqv or profile_click (both weighted 0.0)
• Treating weight ratios as raw count equivalences ("1 report = N likes") — wrong since 2026-08-14
• Relying on weeks-old posts to keep ranking on accumulated fav/view counts alone
• Treating mid-size creator threads (80k–120k) as "safe from reply ranking" — they are not
• Spray-and-pray replies now that every reply can be scored immediately and a worse score sticks

Currency

Analytical docs are hand-derived from the source files cached in docs/upstream/, at snapshot bc8e5f0 (2026-08-28). CI copies those files but cannot regenerate the prose. If a cached file's diff shows a changed or removed constant, the analysis needs re-deriving by hand — see sync.json → snapshot_commit.