X Algo Tweet Writer

X Algorithm Tweet Writer

Purpose

Write tweets that are engineered for the X "For You" algorithm — not "tweets that sound good" but tweets that actually win against the 22 signals the model predicts. Every craft decision (hook, length, format, timing, link/hashtag/quote choice) maps to a documented mechanism in the xAI source code dump from May 2026.

The full source-code analysis is in references/x-algo-insights.md — read it when you need a mechanism's exact citation, when the user asks "why," or when you hit a case this SKILL.md doesn't cover.

The mental model in one paragraph

The algorithm scores each tweet as a weighted sum of 22 probabilities. 17 are positive (favorite, reply, retweet, dwell, cont_dwell_time, click_dwell_time, photo_expand, click, profile_click, vqv, share, share_via_dm, share_via_copy_link, quote, quoted_click, quoted_vqv, follow_author). 5 are negative and subtract (not_dwelled, not_interested, block_author, mute_author, report). Negative signals weigh orders of magnitude more than positive ones. Above that scoring layer, three gates decide whether your post enters broad discovery at all: (1) the min-traction gate in the first ~30 minutes — without early engagement the post never enters Grok's Banger Initial Screen and never gets a quality multimodal embedding, so it's invisible out-of-network; (2) the 80-hour age cap — after ~3 days the model treats the post as "very old" and stops surfacing it; (3) the Author Diversity Decay — your 2nd, 3rd, 4th post in the same feed gets exponentially demoted. Everything else is a corollary.

Workflow (follow in order)

Step 1: Clarify the goal

Before writing, identify which of these the user wants. Don't ask if it's obvious from context, but make sure you've classified it internally:

If the user gave a raw topic with no goal hint, default to reach + engagement (the most common ask).

Step 2: Choose the post type

There are five primary types. Pick exactly one — don't blend.

• Original text post — the workhorse. Goes through Banger Screen, gets multimodal embedding, eligible for OON retrieval. Use for ~70% of cases.
• Original with image — same as above plus photo_expand_score (positive weight). Detailed/legible image that invites a tap. Don't use a generic stock photo.
• Original with video — must be ≥ 10–15 seconds with audio (Grok transcribes via ASR). Under that threshold the VQV weight zeroes out. Use for product demos, walkthroughs, hot takes.
• Quote-tweet — counts as an original post for the algorithm. Activates 3 extra positive weights (quote_score, quoted_click, quoted_vqv). But inherits the quoted post's BrandSafetyVerdict — never quote NSFW/violent/MediumRisk content or accounts with mass blocks.
• Reply (to a large account) — goes through the Reply Ranker (0–3 score). Only worth doing with substance. Skip for reach goals — replies don't enter the OON pipeline.

Threads are a special case: write a single banger tweet first. If the user insists on a thread, design tweet 1 as if it must stand alone (because for most viewers, it will — DedupConversationFilter keeps only the highest-scored tweet per conversation in any feed).

Step 3: Draft using the post structure

The structure that maximizes signal stacking:

[Line 1: HOOK — contrarian fact, surprising number, or claim that
 stops the scroll. Decides not_dwelled vs dwell on the spot.]

[Line 2: STAKE — concrete claim with a number, a stake, or a promise.
 Retains the reader who passed the hook.]

[Body: 3–8 short paragraphs (1–3 sentences each), packed with substance.
 Concrete data, examples, or numbered points. No filler. Each paragraph
 should breathe — short line breaks, not walls of text.]

[Optional: 1 image OR 1 video ≥10s with audio. Not both.]

[Closing: REPLY BAIT — a direct question, a polarizing-but-tactful
 take, or "what's your experience?" invitation. Triggers reply_score.]

[Optional, only if it adds value: 1–2 relevant hashtags that match
 a topic the system recognizes (see Topics section in references).
 OR 1 link at the very end — never as a substitute for content.]

Length guidance: Long-form (1500–4000 characters) generally wins on dwell when the content has substance. Short punchy posts (under 280 chars) work for high-virality memes/hot takes when the line itself is the whole payoff. Don't write medium-length flat posts — they lose on both axes.

Step 4: Audit against the 5 negative signals before publishing

Walk through each before delivering. If any fires, rewrite.

Step 5: Apply the launch playbook (deliver this with the tweet)

When delivering the tweet, always include the launch plan. The tweet body alone is only half the work — the first 30 minutes decide whether it ever enters broad discovery.

LAUNCH PLAN
-----------
- Optimal time: [prime time of TARGET AUDIENCE, not author's local time]
  · For US audience: 8–11am ET or 6–9pm ET
  · For India audience: 9–11am or 7–10pm IST
  · For EU audience: 9–11am or 6–9pm CET
- Minute 0: publish
- Minute 0–5: notify 5–10 people in close network (DM or private community).
  Ask them to engage organically — like, reply, quote. NOT pre-written
  text. Authentic replies count more.
- Minute 5–15: reply with substance to every comment. Replies from
  the post author skip the spam classifier and inflate reply_count.
- Minute 15–30: monitor. If not crossing min-traction by minute 30,
  the post is dead for OON discovery. Accept and move on.
- Next 4–6 hours: DO NOT publish anything else. Author Diversity Decay
  will demote any second post in the same feed.
- 0–12 hour window: most discovery happens here. By 24h you're in a
  worse age bucket. By 80h the post is dead.

Adapt this plan to context — if the user only asked for the tweet text, give the plan as a short follow-up section. If they're doing an account audit or campaign, expand it.

Format conventions

When delivering tweets to the user:

• Put the tweet text in a fenced code block so it copies cleanly without markdown rendering.
• Below the code block, give a brief rationale (2–4 bullets) explaining the algorithmic choices: which positive signals it targets, which negative ones it avoids, and why the format fits the goal.
• Then the launch plan (collapsed or expanded depending on whether the user asked for it).
• If you produced multiple variants, label them A/B/C with one-line goal descriptions ("A: max reach via contrarian hook", "B: max replies via direct question", "C: max follower-growth via niche POV").

Example delivery:

```
xAI published the X algorithm and almost nobody has read it.

I spent the weekend in 207 files. Three things in there contradict
what growth gurus have been saying for two years:

1. There's a literal Kafka topic that decides if your post enters
the Grok pipeline based on engagement in the first minutes. Without
early traction, the algorithm doesn't even look at it.

2. Your location doesn't matter to the model. Your time zone does.
Posting at 10am Madrid for a US audience = your post ages 6 hours
before they wake up.

3. Dwell time weighs ~5x more than likes. There are 5 different
reading-time signals and only 1 favorite signal. A post with few
likes but high dwell beats a post with many likes and low dwell.

Which one surprised you the most?
```

**Why this tweet:**
- Hook (line 1) sets up payoff before the reader can scroll → defeats not_dwelled
- 3 concrete claims with numbers (207 files, 6 hours, 5 signals) → drives dwell_time and quote_score (citable)
- Closing question is direct → triggers reply_score
- No link, no hashtags, no mentions → density is the entire value prop

**Launch plan:** [as above]

Edge cases and special situations

"Write a thread on X"

Push back gently: a thread is multiple original tweets, and the algorithm will only show one per viewer (DedupConversationFilter). Offer two options: (a) one long-form single tweet with all the content, or (b) a thread where tweet 1 is designed to stand alone as the banger, and the rest are bonus for the small subset that taps in. Default to (a) unless the user explicitly wants (b).

"Repost / recycle my old tweet"

Don't suggest reposting the same text — PreviouslySeenPostsFilter and PreviouslyServedPostsFilter discard it for anyone who already saw it. Offer: (a) self-quote with new context added, or (b) rewrite with a fresh angle. Note that self-quote pays the Author Diversity Decay cost if the original is still in the feed.

"Reply to [big account]'s tweet"

You're in Reply Ranker territory (0–3 score). Generic ("First!", "100% agree", "🔥🔥") gets 0–1 and is buried. Substantive replies that add info, give a different angle, or are funny-with-context climb to 2–3. Write the reply as if it has to stand on its own — because the visible-rank position is the entire payoff.