An OpenClaw skill that scores and optimizes your tweets based on X's actual open-source ranking algorithm — not generic growth-hacker tips.
We read all ~21,000 lines of X's official open-source recommendation algorithm (released February 2026) and extracted the real ranking rules into a scoring engine your agent can use.
X's "For You" feed is ranked by a Grok-based transformer (codename Phoenix) that predicts 19 engagement actions for every candidate tweet. The final ranking score is:
score = Σ(weight_i × P(action_i))
This skill encodes the structural rules from that pipeline — what actions exist, how they're weighted, what gets filtered, what gets penalized — into a practical scoring system.
Give your agent a draft tweet and get a detailed score:
🐦 Tweet Composer — Score: 84/100
✅ Length: 142 chars (sweet spot 100-200)
✅ No links in body
✅ Native image attached (+P(photo_expand) boost)
✅ Ends with question (drives P(reply))
⚠️ No video (missing P(video_quality_view) signal)
❌ Posted at 22:00 CET (off-peak for EU audience)
📊 Predicted Action Boost:
├─ P(reply): HIGH — question drives discussion
├─ P(favorite): HIGH — visual + clear value
├─ P(share_via_dm): MEDIUM — niche but shareable
├─ P(dwell): HIGH — image makes people stop
└─ P(not_interested): LOW ✅
💡 Suggestions:
→ Post at 16:00 CET for peak EU engagement
→ Consider a short video walkthrough instead of screenshot
✏️ Optimized version:
"Built a doctor for my Mac. One command: health score,
security audit, cleanup suggestions. Open source.
What would you add to the checklist? 🩺"
→ Reply with: GitHub link
The algorithm's DedupConversationFilter keeps only the highest-scored tweet per conversation. The skill ensures your first tweet is the strongest hook.
For each draft, the skill estimates impact on all 19 predicted actions:
- 15 positive (favorite, reply, repost, quote, share, share_via_dm, share_via_copy_link, click, profile_click, video_quality_view, photo_expand, dwell, dwell_time, quoted_click, follow_author)
- 4 negative (not_interested, block_author, mute_author, report)
Built-in content mix guidance based on empirical analysis:
- 40% entertaining, 30% educational, 20% inspirational, 10% promotional
Discoveries from reading the actual source code:
| Finding | Details |
|---|---|
| 19 actions predicted simultaneously — not just "engagement" | Each tweet gets 19 separate probability scores |
| Author Diversity Scorer — your 2nd tweet gets ~55% score, 3rd ~33% | Exponential decay per author in a single feed |
| Candidate Isolation — each tweet scored independently | Candidates can't see each other in the attention mask |
| Video duration gate — short clips don't get VQV weight | Video must exceed minimum duration threshold |
| Deep reply chains filtered — only first-level replies survive | In-network store drops reply-to-reply-to-reply |
| Share via DM has its own weight — separate from generic share | 3 separate share signals: generic, DM, copy link |
| dwell_time is continuous — not boolean, measured in seconds | Longer reading time = proportionally higher score |
| Signed action encoding — model knows what you did AND didn't do | Actions encoded as +1 (did) / -1 (didn't) |
clawhub install tweet-composergit clone https://github.com/minilozio/tweet-composer-skill.git
# Copy to your OpenClaw skills directorytweet-composer-skill/
├── SKILL.md # Main skill instructions + scoring rubric
├── references/
│ └── algorithm-rules.md # Complete algorithm rules engine (19 actions, filters, scorers)
├── assets/
│ └── banner.svg
├── LICENSE
└── README.md
"Score this tweet: Just shipped a new feature. Check it out!"
"Optimize this thread for maximum reach"
"What's wrong with this tweet? Why did it get low engagement?"
"Write me a tweet about [topic] that maximizes reply potential"
"Should I post this now or wait for peak hours?"
All rules are derived from X's official open-source recommendation algorithm, released February 2026. We read every file — ~21K lines of Rust + Python/JAX.
MIT — see LICENSE
Built by @minilozio 🦎