X Boss

Read X's own ranking code. Let your AI read your account. Post what travels.

X's For You algorithm, made usable. X open-sourced the code that ranks the For You feed (xai-org/x-algorithm, Apache-2.0; the August 2026 release added the real production weights, filters and safety rules). This kit turns that code into things you can actually use:

• Ask your AI to read your account. Point Claude / ChatGPT / Cursor / Claude Code at this repo and say "analyze my X account". It asks for your handle, gets your posts (a script you run in your own browser; nothing leaves your machine), and tells you, in code terms, why your posts don't travel and what to post next, drafted in your own voice.
• The weights as data, auto-synced. data/weights.json mirrors the production parameters with a file:line for every number; a GitHub Action re-extracts them daily and opens an issue whenever X changes one. Watch this repo to hear about it first.
• A scorer that reproduces the ranking arithmetic, a draft linter with code-derived rules, an interactive simulator, and a knowledge base with every mechanism cited.

Nothing here predicts impressions. The model's per-viewer probabilities are not public. What the code does make clear is which levers exist, what gates a post out entirely, and what quietly caps reach.

Ask your AI (60 seconds)

Claude Code / Cursor / any agent that can read a repo

git clone https://github.com/gargvr/xboss && cd xboss

then say: "Use the xboss skill in this repo to analyze my X account and draft my next week of posts." (Claude Code discovers .claude/skills/xboss/SKILL.md automatically.) The agent follows skills/xboss/SKILL.md: it asks for your handle, has you run scripts/collect_profile.js on your profile (DevTools console, copies your posts as JSON to your clipboard, local only), runs python -m xboss analyze me.json, explains the result against knowledge/, and lints every draft it writes.

Claude Desktop / MCP clients

pip install "mcp[cli]"
# claude_desktop_config.json → {"mcpServers": {"xboss": {"command": "python", "args": ["/ABS/PATH/xboss/mcp/server.py"]}}}
# Claude Code: claude mcp add xboss -- python /ABS/PATH/xboss/mcp/server.py

Tools: get_weights, score_post, lint_draft, analyze_account, cold_start, read_knowledge.

ChatGPT / Claude web (no repo access)

Paste prompts/system_prompt.md into a new chat, then paste your exported posts. Same procedure, no tools.

Rate limits and your account (read before exporting)

X throttles its own web endpoints per account in roughly 15-minute windows and, per the upstream bdsm/ component, scores automation-like behaviour. scripts/collect_profile.js therefore scrolls at a human pace with random pauses, stops at ~40 posts, and performs no actions. Run it once, yourself, in a foreground tab. Don't let an AI agent loop it against your logged-in session, and if you see "Rate limit exceeded" or "Something went wrong", wait 15-30 minutes instead of retrying. Reading your own profile is normal use; scripted repetition is what gets throttled.

What the account read tells you

python -m xboss analyze me.json prints, for example:

25 followers · 143 following · no Premium · joined 2026-05 · 31 posts (30 originals)
median views, originals            51        originals with zero likes    57%
median views, media originals      77        median views, text-only      40
cold-start window (≤1000 followers) yes      Under the Hood available     not yet

- Median original gets 51 views against 25 followers (2.1x): roughly the follower graph plus profile visits.
  In code terms this is a candidate-generation problem: posts that never enter the out-of-network index
  (original + safety-clean + ≥1 like, evicted at 24h) are invisible to everyone else. [phoenixRankAllCandidateProcessor.strato:437-459]
- 57% of originals got zero likes: a post with zero likes never enters the Phoenix retrieval index and builds
  no SimClusters vector (likes only, 8h half-life). The first like is the door.
- Your best originals by engagement: "I open-sourced an AI that runs a faceless Shorts channel…" (198 views, 6 likes) …
  whatever these have in common is your lane; retrieval represents a post as author-hash + content codes.
- At 25 followers you are inside the cold-start window: every For You request lifts one small-account original
  with <1000 views and <24h age to slot ~15, IF it is already a candidate for that viewer. [author_cold_start.rs]

…followed by a plan (lane, first-like protocol, mutual-follow graph, cadence, what to stop) with a file reference for each item.

The weights (auto-synced)

Mirrored from upstream commit 9b0dc31 (2026-09-04); param.rs header says last production sync 2026-09-03T16:23:24Z; extracted 2026-09-04. Weights multiply the viewer's predicted probability of the action, not raw counts.

action	param	weight	where
share via copy link	ShareViaCopyLinkWeight	20	home-mixer/params/param.rs:357
reply	ReplyWeight	5	home-mixer/params/param.rs:315
quote	QuoteWeight	5	home-mixer/params/param.rs:364
share via DM	ShareViaDmWeight	5	home-mixer/params/param.rs:351
follow author	FollowAuthorWeight	4	home-mixer/params/param.rs:377
share	ShareWeight	2	home-mixer/params/param.rs:350
repost	RetweetWeight	1	home-mixer/params/param.rs:328
like	FavoriteWeight	0.5	home-mixer/params/param.rs:314
click (open post)	ClickWeight	0.4	home-mixer/params/param.rs:341
open link	OpenLinkWeight	0.2	home-mixer/params/param.rs:342
video quality view (>10 s, viewer <10k followers)	VqvWeight	0	home-mixer/params/param.rs:349
photo expand	PhotoExpandWeight	0.05	home-mixer/params/param.rs:329
video open	VideoOpenWeight	0.07	home-mixer/params/param.rs:335
quoted click	QuotedClickWeight	0.05	home-mixer/params/param.rs:365
quoted video quality view	QuotedVqvWeight	0	home-mixer/params/param.rs:371
profile click	ProfileClickWeight	0	home-mixer/params/param.rs:343
dwell (binary)	DwellWeight	0.05	home-mixer/params/param.rs:363
dwell time (per predicted second)	ContDwellTimeWeight	0.004	home-mixer/params/param.rs:407
click dwell time (per second)	ContClickDwellTimeWeight	0	home-mixer/params/param.rs:413
active secs residual	ContActiveSecs5mResidualNormWeight	0	home-mixer/params/param.rs:449
post unexplored (in-network only)	PostUnexploredWeight	0.02	home-mixer/params/param.rs:383
reply boost when author and viewer follow each other (originals only)	BidirectionalFollowReplyWeightBoost	15	home-mixer/params/param.rs:316
dwell boost, mutual follow	BidirectionalFollowDwellWeightBoost	0	home-mixer/params/param.rs:322
report	ReportWeight	-234	home-mixer/params/param.rs:474
mute author	MuteAuthorWeight	-58.8	home-mixer/params/param.rs:468
not interested	NotInterestedWeight	-43.2	home-mixer/params/param.rs:456
block author	BlockAuthorWeight	-31.2	home-mixer/params/param.rs:462
not dwelled	NotDwelledWeight	-0.02	home-mixer/params/param.rs:475

After the sum: out-of-network ×0.75 (also applied to replies/reposts from followed accounts: True); author diversity decay 0.5, floor 0.25 (2nd post ×0.625); cold-start lift for authors ≤ 1000 followers with < 1000 views and < 24h age, to slot 15; DPP reranker theta 0.65 over top 150; posts older than 48h never enter.

Full parameter set: data/params.json (189 params), data/constants.json, data/retention_windows.json. Every change lands in CHANGELOG.md and as an issue labelled upstream-change.

Ten things the code says (each cited in knowledge/)

1. Score = Σ weight × that viewer's predicted probability of each action. Copy-link share 20, reply 5 (20 for mutual follows on originals), quote 5, DM share 5, follow 4, like 0.5; report −234, mute −58.8, not-interested −43.2. Weights multiply probabilities, not counts.
2. To be recommended to non-followers a post must be an original, pass safety filtering, and get ≥1 like; it leaves the retrieval index at 24 h regardless. Ranking's cap is 48 h.
3. The shipped model does not consume engagement counts, verified/Premium, raw text, or links; content enters as semantic IDs from a multimodal embedding. Follower count: not consumed either, but since the Aug 17 commit the mixer sends it to the model (the sync bot caught that on day one; see CHANGELOG.md).
4. ≤1,000 followers? One original per request with <1,000 views and <24 h age is lifted to slot ~15, if it is already a candidate.
5. Only post 1 of a thread can leave your follower graph. OON replies and reposts are removed before scoring; quotes are originals.
6. Your 2nd post in a viewer's pool is ×0.625, 3rd ×0.44 (floor 0.25). Replies/reposts to your own followers are ×0.75.
7. A DPP reranker (theta 0.65) keeps ~50 of the top 150, dropping near-duplicate takes. Same topic, different angle, or skip.
8. SimClusters counts likes only, halves every 8 h, needs ≥8 likes to seed others: who likes you first matters more than how many.
9. No link penalty exists; URL reputation does: a low-quality redirect chain hides the post from non-followers; a pinned post with a bad link hides your whole account from non-followers for a week; engagement bait is labelled regardless of reputation.
10. Reputation (PageRank ≥54) is a shield against automatic spam labels, not a ranking boost. Premium has no ranking term.

Myths checked against the code: knowledge/04-myths.md. Interactive simulator + full manual: docs/index.html (GitHub Pages).

CLI

pip install -e .            # optional; or just run python -m xboss from the repo root (no dependencies)

xboss weights                          # mirrored weights with file:line refs
xboss explain ReplyWeight              # where a parameter lives upstream (with GitHub link)
xboss score --probs '{"favorite":0.03,"reply":0.002,"share_via_copy_link":0.0005,"dwell_time_secs":6,"not_dwelled":0.6}' [--mutual --oon --k 1]
xboss lint "draft text" [--media --premium --reply --thread]
xboss analyze me.json [--csv content.csv] [--paste profile.txt] [--handle h --followers n --joined 2026-05]
xboss cold-start --followers 25 --views 40 --age-hours 3
python scripts/sync.py --clone          # re-extract from upstream; exit code 3 = something changed
python -m pytest -q

Repository map

data/               weights.json · params.json · constants.json · retention_windows.json  (auto-synced)
xboss/             score.py (RankingScorer arithmetic) · lint.py · analyze.py · intake.py · cli.py
knowledge/          01-scoring · 02-reach · 03-suppression · 04-myths · 05-playbook   (every claim cited)
skills/xboss/      SKILL.md   the AI procedure (Claude Code skill format; also copied to ./SKILL.md)
AGENTS.md llms.txt  instructions/index for coding agents
prompts/            system_prompt.md   paste-in version
mcp/                server.py   MCP tools
scripts/            collect_profile.js (export your posts, locally) · sync.py (extract+diff) · render_readme.py
docs/               index.html   simulator + field manual (GitHub Pages)
.github/workflows/  sync.yml   daily upstream sync → commit → issue on change

Honesty notes

• Values are mirrored from the upstream repo's own "production defaults" (param.rs header carries the sync timestamp). X runs experiments; not every live variant is in the repo.
• Withheld upstream: Grox LLM prompts, URL-reputation verdict logic, BDSM thresholds, the duplicate-text SQL, "some botmaker rules". The kit says "not in the repo" where that matters.
• The linter's engagement-bait check is a phrase heuristic standing in for a withheld LLM classifier; it is labelled heuristic in every finding.
• Analysis runs locally on data you exported from your own profile. The kit never logs in, posts, follows, or sends anything.

Contributing

Issues and PRs welcome: new intake loaders (other analytics exports), better voice-matching prompts, translations of the knowledge base, corrections with a file:line. Run python -m pytest -q before opening a PR.

License

MIT for this kit. The upstream code it reads is Apache-2.0 © X Corp / xAI; this project is not affiliated with X.