---

name: x-algo description: "Score and optimize tweets against the X algorithm's 19 engagement signals. Use for: optimize tweet, make viral, tweet score, algorithm check, rate tweet, tweet formats."

X Algorithm Tweet Optimization

Analyze and optimize tweets based on the open-source X ranking algorithm (xai-org/x-algorithm). This skill encodes how the "For You" feed actually works — the 19 engagement signals, weighted scoring, and ranking factors — into actionable guidance for crafting tweets that perform.

For full algorithm details: read references/x-algorithm.md.

When to Use This Skill

Activate this skill when the user:

• Asks to optimize, score, rate, or review a tweet or draft
• Wants to craft a viral tweet or asks "make this go viral"
• Asks about the X / Twitter algorithm or how the "For You" feed works
• Wants to compare tweet drafts or A/B test versions
• Asks about tweet formats or what format to use for a topic
• Says anything like: "optimize tweet", "tweet score", "algorithm check", "rate this tweet"

Do NOT activate for:

• General social media strategy (this is tweet-level, not account-level)
• Non-X platforms (Instagram, LinkedIn, TikTok)
• Scheduling, posting, or automation questions

CLI Tool

All commands run from this skill directory:

cd <skill-directory>

Analyze

bun run x-algo.ts analyze "<tweet text>"

Full analysis: composite score, signal breakdown, format detection, hook analysis, suggestions, and warnings. Use this for comprehensive feedback on a draft tweet.

Score

bun run x-algo.ts score "<tweet text>"

Quick one-line score (0-100) with format and hook summary. Use for fast iteration on multiple drafts.

Signals

bun run x-algo.ts signals "<tweet text>"

Detailed breakdown of all 19 algorithm signals (0-10 scale each). Shows which positive signals the tweet triggers and any negative signal risks.

Compare

bun run x-algo.ts compare "tweet A|||tweet B|||tweet C"

Score multiple drafts in one call, ranked by score. Separate tweets with |||. Use for A/B testing drafts — an agent can generate several versions and pick the winner. Supports all flags (--image, --video, --reply, --json).

Formats

bun run x-algo.ts formats

Lists high-performing tweet format templates with examples: thread starter, hot take, data drop, story, listicle, question, contrarian reframe, before/after.

Reference

bun run x-algo.ts reference [topic]

Quick lookup on algorithm topics: signals, scoring, filters, diversity, retrieval, etc.

Media Flags

All analysis commands (analyze, score, signals) support media flags:

bun run x-algo.ts analyze "<tweet text>" --image
bun run x-algo.ts analyze "<tweet text>" --video
bun run x-algo.ts analyze "<tweet text>" --image --video

Use --image when the tweet will include an attached image (screenshot, chart, meme, etc.). Boosts photo_expand, dwell, click, and favorite signals.

Use --video when the tweet will include an attached video (demo, walkthrough, clip, etc.). Unlocks the video_quality_view signal (otherwise stuck at 0) and boosts dwell, share, and favorite signals.

When crafting tweets, always consider whether media should be part of the post — then pass the appropriate flag to get an accurate score.

Reply Flag

All analysis commands (analyze, score, signals, compare) support the --reply flag:

bun run x-algo.ts analyze "<tweet text>" --reply
bun run x-algo.ts score "<tweet text>" --reply
bun run x-algo.ts compare "reply A|||reply B" --reply

Use --reply when the tweet is a reply to someone else's post. The X algorithm applies a 0.75x discount to ALL reply tweets (ReplyDownRankerParams.replyDownRankFactor). With --reply, the tool:

• Applies the 0.75x discount to the composite score
• Shows both the discounted score and the raw (undiscounted) score
• Generates reply-specific suggestions for overcoming the discount

When an agent is building conversation (the 75.0x signal), it needs to reply — but those replies face a 25% penalty. Use --reply to ensure reply content is strong enough to still rank well despite the discount.

All commands also support --json for structured output.

How the X Algorithm Works (Summary)

The "For You" feed is ranked by a multi-task neural network (MaskNet architecture). For every tweet-user pair, the model predicts engagement probabilities. The final score is a weighted sum.

Published Production Weights (twitter/the-algorithm, April 2023)

These are the actual weights from the open-source release — they determine how much each engagement type affects ranking:

| Model Head | Weight | What It Means | |-----------|--------|---------------| | Reply Engaged by Author | 75.0 | Someone replies AND author engages back — by far the strongest signal | | Reply | 13.5 | Will they write a reply? | | Good Profile Click | 12.0 | Will they visit profile AND engage? | | Good Click V1 (convo+fav/reply) | 11.0 | Will they click into convo and like/reply? | | Good Click V2 (convo+2min dwell) | 10.0 | Will they click in and stay 2+ minutes? | | Retweet | 1.0 | Will they retweet? | | Like | 0.5 | Will they tap the heart? | | Video 50% Playback | 0.005 | Will they watch 50%+ of video? | | Negative Feedback | -74.0 | Will they click "show less", block, or mute? | | Report | -369.0 | Will they report the tweet? Heaviest penalty |

Key insight: Likes (0.5x) and retweets (1.0x) barely matter. Replies (13.5x) and especially conversation engagement (75.0x) dominate. A tweet that sparks genuine back-and-forth is worth 150x more than one that just gets likes.

The 19 Display Signals

Our tool maps to 19 granular signals (0-10 scale each) for detailed analysis, based on the model heads above plus signals appearing in the training data (share, share_dm, share_copy_link, photo_expand, bookmark, follow, dwell).

Ranking Multipliers (Applied After Scoring)

• Out-of-Network Discount: 0.75x for tweets from accounts you don't follow
• Reply Discount: 0.75x for reply tweets
• Author Diversity: Exponential decay — 2nd post from same author gets ~0.625x, 3rd ~0.44x, approaches 0.25x floor
• Feedback Fatigue: "See Fewer" reduces score to 0.2x, recovering over 140 days
• Freshness: AgeFilter removes tweets past maximum age
• Original Content: Retweet deduplication means original posts beat reposts

Viral Tweet Methodology

When the user asks you to craft or optimize a tweet, follow this process:

1. Understand the Goal

Before writing, clarify:

• What's the core message or insight?
• Who is the target audience?
• What action do you want from readers? (engagement, shares, follows)

2. Choose the Right Format

Select a format that matches the content. Run bun run x-algo.ts formats for templates.

| Content Type | Best Format | Key Algorithm Signals | |-------------|-------------|----------------------| | Opinion | Hot take, contrarian reframe | reply (13.5x), conversation (75.0x) | | Question | Question/poll | reply (13.5x), conversation (75.0x), click (11.0x) | | Experience | Story, before/after | profile_click (12.0x), dwell (10.0x) | | Knowledge/insight | Data drop, listicle | click (11.0x), dwell (10.0x), share | | Announcement | Thread starter | click (11.0x), dwell (10.0x) |

3. Craft the Hook

The first line determines whether someone stops scrolling. Strong hooks:

• Question: "What's the most underrated skill in tech?"
• Bold claim: "Unpopular opinion: consistency is overrated."
• Data point: "Our product went from 0 to $1M ARR in 6 months."
• Story opener: "I almost got fired last week."
• Listicle: "7 things nobody tells you about fundraising:"

Weak hooks: generic statements, "I think...", "Just wanted to share...", starting with a link.

4. Maximize Positive Signals (Ranked by Algorithm Weight)

Focus efforts on the highest-weight signals first:

Conversation (75.0x) — THE most important factor. Write tweets that spark genuine back-and-forth. Ask questions, invite discussion ("what do you think?", "agree or disagree?"), present debatable ideas. A tweet that generates thoughtful replies where you engage back is worth 150x more than one that just gets likes.

Replies (13.5x) — Ask genuine questions. Present something debatable. Leave room for people to add their take. Use conversation starters like "what's your take?", "thoughts?", "am I wrong?"

Profile Click + Engage (12.0x) — Demonstrate unique expertise. Share personal experience, original frameworks, build logs. Make people think "who is this person?" and click through to your profile. Use patterns like "I built...", "after X years...", "my approach to...".

Click Into Convo (11.0x) — Tease information. Thread starters. "Here's what I found:". Content that creates curiosity and makes people click to read more.

Dwell / Deep Engagement (10.0x) — Substance. Multi-line formatting. Storytelling. Content that rewards reading carefully. Longer tweets with line breaks.

Reposts (1.0x) — Content that makes the sharer look smart. Data, frameworks, curated knowledge. Relatively low weight.

Likes (0.5x) — Clear, quotable insight. But likes barely matter in the algorithm — focus on replies and conversation instead.

Video (0.005x) — Nearly irrelevant weight. Don't optimize for video views — the algorithm barely cares.

5. Minimize Negative Signals

What triggers each negative signal:

Negative Feedback (-74.0x) — Generic content, engagement bait ("like if you agree"), excessive hashtags, aggressive tone, excessive self-promotion, excessive caps. The model combines "show less", block, and mute into one prediction head. Getting any of these is 148x worse than getting a like.

Report (-369.0x) — Spam, misleading content, harassment, policy violations. The single heaviest penalty in the entire algorithm — 738x the weight of a like. Getting reported is algorithmic death.

6. Analyze and Iterate

Run the draft through the CLI:

bun run x-algo.ts analyze "your draft tweet"
bun run x-algo.ts analyze "your draft tweet" --image   # if attaching an image
bun run x-algo.ts analyze "your draft tweet" --video   # if attaching a video
bun run x-algo.ts analyze "your reply tweet" --reply   # if replying to someone

Review the score, signal breakdown, and suggestions. Iterate on the draft. Target:

• Score: 65+ is good, 85+ is great, 95+ is exceptional
• For replies (with --reply): 55+ is good, 70+ is great (raw score before 0.75x discount)
• No negative signal above 3/10
• Reply signal at 7+/10 (highest-impact signal in the algorithm)
• At least 2-3 other positive signals at 6+/10

Compare multiple versions in one call:

bun run x-algo.ts compare "version A|||version B|||version C"
bun run x-algo.ts compare "version A|||version B" --image
bun run x-algo.ts compare "reply A|||reply B" --reply

7. Final Checklist

Before posting, verify:

• [ ] Strong hook (first line grabs attention)
• [ ] Under 280 characters (or intentionally long with line breaks)
• [ ] At least one engagement trigger (question, data, opinion, story)
• [ ] No engagement bait patterns
• [ ] No excessive hashtags (max 2-3)
• [ ] No spam-like CTAs
• [ ] Consider adding media (image, video, screenshot) — use --image or --video flag when scoring

Anti-Patterns to Avoid

These patterns hurt your algorithm ranking:

1. Engagement bait — "Like if you agree", "RT for reach", "Follow for follow". Drives not_interested and mute signals.

2. Excessive hashtags — More than 3 hashtags reads as spam. The algorithm doesn't use hashtags for ranking; it uses content understanding.

3. Thread padding — Starting a 1-tweet thought as a "thread" just for engagement. Only thread when you have multiple substantive points.

4. All caps — Reads as shouting. Increases mute signals.

5. Over-posting — Author Diversity Scorer applies exponential decay. Space posts 2+ hours apart for maximum individual reach.

6. Generic takes — "AI is changing the world" with no specific insight. Low engagement = low ranking.

7. Link-only tweets — A URL with no context. Low dwell time, low engagement signals.

8. Crypto spam patterns — Airdrop, whitelist, giveaway language triggers spam filters and user mutes.

Edge Cases

• Thread tweets: Score the first tweet as the hook — it determines whether people click into the thread. Subsequent tweets don't need standalone hooks.
• Link-only tweets: Warn the user these score poorly (low dwell, low engagement). Always suggest adding context or a hook above the link.
• Very short tweets (<30 chars): May score low on dwell but can still perform well on reply/conversation signals if the content is provocative or asks a question.
• Tweets about media without flags: If the tweet text references an image or video ("check out this screenshot", "watch this"), remind the user to pass --image or --video for accurate scoring.
• Non-English content: Pattern matching is English-focused. Scores may be less accurate for other languages — note this limitation when relevant.
• Reply to own tweet (self-thread): Use --reply only for replies to other people's posts. Self-threads don't get the 0.75x discount.

File Structure

x-algo-skill/
├── SKILL.md           (this file)
├── x-algo.ts          (CLI entry point)
├── lib/
│   ├── analyze.ts     (tweet analysis engine — 19 signal scoring)
│   └── format.ts      (output formatters)
├── data/
│   └── .gitkeep
└── references/
    └── x-algorithm.md (distilled X algorithm reference)