Claude Code skills for growing on X — built directly on the open-source xai-org/x-algorithm code.
In January 2026 X open-sourced its recommendation algorithm. In May 2026 it shipped the largest update yet — a runnable pipeline, the Grok-based Phoenix ranker, and a content- understanding layer. Most "X growth tips" online still quote the retired 2023 Twitter algorithm.
This repo is different. Every skill here is grounded in what the current code actually does. They turn Claude Code into a co-pilot that audits, writes, and reverse-engineers X posts the way the algorithm really ranks them.
| Skill | What it does |
|---|---|
/x-audit |
Audits a draft post or thread before you publish — PASS / FLAG / BLOCK across 8 algorithm-derived dimensions, plus a rewrite. |
/x-hook |
Writes and ranks scroll-stopping first lines using 7 hook archetypes. |
/x-thread |
Turns notes or an article into a thread with a first post that can stand alone. |
/x-teardown |
Reverse-engineers why a post went viral (or flopped), mechanism by mechanism, and gives you a reusable template. |
/x-reply |
Builds a reply-based cold-start strategy for new accounts and drafts high-quality replies. |
/x-profile |
Audits your handle, bio, header, and pinned post for follow conversion. |
All six are open-core utilities — they make a single thing you're about to do better. The full "tell me exactly what to post every day for 14 days" system is a product; see below.
The 2026 algorithm works on one principle:
Phoenix — a transformer built on xAI's Grok architecture — reads the viewer's engagement history and predicts how likely that specific person is to like, reply, share, follow, or ignore your post. Your post is scored once per viewer. There is no global "virality score."
That single change makes most legacy advice wrong:
- The famous fixed weights ("a reply is worth 13.5× a like") no longer exist — Phoenix has no hand-engineered features, and X does not publish the new numeric weights.
- Out-of-network reach is embedding similarity — entity-dense, single-topic posts get retrieved into the right feeds; vague posts reach no one.
share_via_dmandfollow_authorare predicted actions — "would someone DM this" and "would a stranger follow you from this post" are direct ranking inputs.- Engagement-bait is predicted as
not_interested/reportand scores negative.
The full breakdown — pipeline, the 19 predicted actions, the two hard-coded scoring rules,
and an honest list of what's still unknown — is in
reference/ALGORITHM.md. Every skill reads from it.
/plugin marketplace add lxyang20131208-star/x-algorithm-skills
/plugin install x-algorithm-skills
Copy the skill folders into your skills directory:
git clone https://github.com/lxyang20131208-star/x-algorithm-skills
cp -r x-algorithm-skills/skills/* ~/.claude/skills/Then restart Claude Code. The skills appear as /x-audit, /x-hook, etc.
Just describe what you need — the skills trigger on intent, or invoke them by name:
/x-audit
Here's my draft: "Excited to announce we just launched our new AI tool! Check it out 👇"
/x-teardown
Why did this blow up? <paste a viral tweet>
I'm at 40 followers and posting into the void — help
→ triggers /x-reply
Make this into a thread: <paste your notes>
→ triggers /x-thread
The skills will ask for what they need (your niche, real numbers, the draft). They never invent metrics or fake X handles — if a post is weak because it lacks specifics, they tell you to add real ones.
- Quote fake weight numbers. The 19 predicted actions are public; their weights are not. Anyone selling you exact multipliers for 2026 is guessing.
- Promise virality. The algorithm rewards posts a transformer predicts a real person will act on. These skills stack the odds; they don't manufacture luck.
- Fabricate your results. Specificity is the #1 reach lever — and it has to be true.
These skills are the open-source layer of GrowthHunt — tools for AI founders who have to do their own go-to-market.
If you want the layer above these skills — a personalized 14-day launch calendar that tells you what to post each day, with copy-paste templates from real AI startups that got early traction — that's ViralX, at growthhunt.ai.
x-algorithm-skills improves any single post. ViralX runs the whole launch.
Issues and PRs welcome — especially A/B test results. The algorithm gives us the shape of what matters, not the magnitudes; real experiments are how we close the gap. If you test a hypothesis (hook archetype, link placement, format), share the data.
Not affiliated with X Corp or xAI. Built by reading the public Apache-2.0 source at
xai-org/x-algorithm. The algorithm changes; this repo tracks it on a best-effort basis.
MIT — see LICENSE.