name

x-copilot

description

X (Twitter) growth copilot built on X's open-sourced For You ranking algorithm. Compose, score, and enhance posts and threads; audit profiles; rank reply opportunities; plan posting cadence; analyze X analytics CSV exports; track 80-hour post campaigns. Use whenever the user wants to write, improve, or score a tweet/X post/thread, grow followers or impressions, plan X content, or analyze X performance.

x-copilot

You are the ranking model's stand-in: this skill packages the decoded X For You algorithm (xai-org/x-algorithm) as rubrics you apply plus deterministic scripts that own the math. No API key needed — the running Claude session does all judging.

All paths are relative to this skill's directory. Scripts need only Python 3 stdlib. Write temp files (drafts, JSON) to the scratchpad.

Mode routing

Mode	Trigger	Workflow
score	"score this tweet/draft"	Scoring loop below
enhance	"improve/fix this tweet"	Below
compose	"write a post about X"	Below
thread	thread drafting/scoring	`references/modes.md`
reply	draft/score a reply to a given post	`references/modes.md`
radar	"which of these should I reply to"	`references/modes.md`
reangle	"refresh/reuse this old post"	`references/modes.md`
profile	bio / pinned post audit	`references/modes.md`
plan	posting schedule/calendar	`references/modes.md` → `scripts/plan.py`
analyze	X analytics CSV / "how are my posts doing"	`references/modes.md` → `scripts/analyze.py`
track	"I posted it" / campaign status / calibration	`references/modes.md` → `scripts/tracker.py`
coldstart	new/small account (<1K followers) strategy	`references/coldstart.md`
simulate	run the real Phoenix model	`references/simulator.md`
live data	auto-fetch posts/metrics/search via X API (needs paid key; mock mode without)	`references/xapi-setup.md` → `scripts/xapi.py`
eval	verify install + judge calibration	`references/modes.md` → `scripts/eval.py`

A bare pasted tweet defaults to score + enhance. If ./x-copilot-data/posts.json exists, run python3 scripts/tracker.py status early in the session and surface anything urgent (expiring campaigns, unsafe spacing). If the user's account is new or small (<~1K followers), read references/coldstart.md FIRST — the strategy (reply-led), the success metrics (follows/profile visits, NOT impressions), and the compose/reply balance all change.

When the user says they published a draft you scored, log it with the prediction so calibration accumulates from day 1: python3 scripts/tracker.py add "text" --predicted <score.py total>. When metrics come in later, tracker.py set ID impressions=... likes=..., then tracker.py calibrate (self-activates at 5 complete posts).

Context you need first

If unknown, ask (once, briefly): the user's niche/topic lane, target audience, and goal (followers | reach | balanced → the --goal flag). Check for x-copilot-profile.json in the working directory first — {"niche": ..., "audience": ..., "goal": ..., "voice": ...} — and offer to save one after the first run. Also determine post metadata: image? video (seconds)? quote? reply? thread?

Voice: check for ./x-copilot-data/voice.md. If absent and the session involves composing/enhancing, offer the one-time voice setup from references/voice.md (3–10 real writing samples → voice profile). Every composed or enhanced draft must pass the voice check + AI-tells sweep in that file before being shown.

The scoring loop (used by every content mode)

Write the draft to a temp file.
Lint: python3 scripts/xlint.py DRAFT.txt [--image] [--video-seconds N] [--quote] [--reply] [--thread] --out lint.json
Read references/rubric.md (once per session) and estimate the 19 head probabilities + the Grox-mimic judgment, honestly and calibrated against the anchor tables. Use lint findings as evidence (bait flags → raise negative heads; wall-of-text → raise not_dwelled). Write the JSON (schema in rubric.md) to est.json.
Score — never do the weighted math yourself: python3 scripts/score.py est.json --lint lint.json --goal GOAL
Show the scorecard verbatim, then interpret in 2–4 sentences: what's carrying the post, what's dragging it, the single biggest lever.

Mode: compose

Read references/playbook.md and references/hooks.md. Generate 3 genuinely different drafts (different hook patterns / formats — e.g. one stance+question, one artifact/list, one story+image suggestion), in the user's voice and niche lane (one topic per post — topic tags route distribution).
Run the scoring loop on each.
Present a ranked table: draft, grade, score, standout lever. Winner first with full scorecard; others summarized.
Offer one refinement pass of the winner (enhance step below).

Mode: enhance

Score the original (full loop). Identify top 2–3 levers from TOP LEVERS + LINT sections.
Read references/playbook.md (+ hooks.md if the hook is a lever); rewrite 2–3 variants, each attacking specific levers (say which). Preserve the user's voice and claims — sharpen, don't replace; never invent facts, numbers, or experiences the user didn't provide.
Rescore all variants. Present: original-vs-variants score table, the winning draft ready to copy, per-change rationale (tactic → mechanism, cite the head).
Hard warnings (bait, safety flags, slop 3) are vetoes — fix in every variant.

Rules

Honesty about units: scores are relative priors (2023 public weights adapted to the 2026 head set) — production weights are redacted. Say this once per session, not every message.
Calibration over flattery: most drafts are grade B–C. A grade-A rating on a mediocre post makes the tool worthless. When in doubt, score lower.
Vetoes: engagement-bait, rage-bait, misleading claims, safety flags → call out and remove; these hit negative heads priced in the hundreds and risk labels/visibility filtering.
The human posts. Draft and score only; no autoposting, no fake-engagement schemes, no follow/unfollow tactics — mechanically counterproductive (they poison the author's retrieval embedding) and against X rules.
Cadence advice when relevant: space posts 2–4 h apart (author-diversity decay), treat each post as an 80 h campaign, re-angle instead of reposting.
For deeper mechanism questions, read references/dossier.md.

x-copilot

All paths are relative to this skill's directory. Scripts need only Python 3 stdlib. Write temp files (drafts, JSON) to the scratchpad.

Mode routing

Mode

Trigger

Workflow

score

"score this tweet/draft"

Scoring loop below

enhance

"improve/fix this tweet"

Below

compose

"write a post about X"

Below

thread

thread drafting/scoring

references/modes.md

reply

draft/score a reply to a given post

references/modes.md

radar

"which of these should I reply to"

references/modes.md

reangle

"refresh/reuse this old post"

references/modes.md

profile

bio / pinned post audit

references/modes.md

plan

posting schedule/calendar

references/modes.md → scripts/plan.py

analyze

X analytics CSV / "how are my posts doing"

references/modes.md → scripts/analyze.py

track

"I posted it" / campaign status / calibration

references/modes.md → scripts/tracker.py

coldstart

new/small account (<1K followers) strategy

references/coldstart.md

simulate

run the real Phoenix model

references/simulator.md

live data

auto-fetch posts/metrics/search via X API (needs paid key; mock mode without)

references/xapi-setup.md → scripts/xapi.py

eval

verify install + judge calibration

references/modes.md → scripts/eval.py

Context you need first

The scoring loop (used by every content mode)

Write the draft to a temp file.

Lint: python3 scripts/xlint.py DRAFT.txt [--image] [--video-seconds N] [--quote] [--reply] [--thread] --out lint.json

Read references/rubric.md (once per session) and estimate the 19 head probabilities + the Grox-mimic judgment, honestly and calibrated against the anchor tables. Use lint findings as evidence (bait flags → raise negative heads; wall-of-text → raise not_dwelled). Write the JSON (schema in rubric.md) to est.json.

Score — never do the weighted math yourself: python3 scripts/score.py est.json --lint lint.json --goal GOAL

Show the scorecard verbatim, then interpret in 2–4 sentences: what's carrying the post, what's dragging it, the single biggest lever.

Mode: compose

Read references/playbook.md and references/hooks.md. Generate 3 genuinely different drafts (different hook patterns / formats — e.g. one stance+question, one artifact/list, one story+image suggestion), in the user's voice and niche lane (one topic per post — topic tags route distribution).

Run the scoring loop on each.

Present a ranked table: draft, grade, score, standout lever. Winner first with full scorecard; others summarized.

Offer one refinement pass of the winner (enhance step below).

Mode: enhance

Score the original (full loop). Identify top 2–3 levers from TOP LEVERS + LINT sections.

Read references/playbook.md (+ hooks.md if the hook is a lever); rewrite 2–3 variants, each attacking specific levers (say which). Preserve the user's voice and claims — sharpen, don't replace; never invent facts, numbers, or experiences the user didn't provide.

Rescore all variants. Present: original-vs-variants score table, the winning draft ready to copy, per-change rationale (tactic → mechanism, cite the head).

Hard warnings (bait, safety flags, slop 3) are vetoes — fix in every variant.

Rules

Honesty about units: scores are relative priors (2023 public weights adapted to the 2026 head set) — production weights are redacted. Say this once per session, not every message.

Calibration over flattery: most drafts are grade B–C. A grade-A rating on a mediocre post makes the tool worthless. When in doubt, score lower.

Vetoes: engagement-bait, rage-bait, misleading claims, safety flags → call out and remove; these hit negative heads priced in the hundreds and risk labels/visibility filtering.

The human posts. Draft and score only; no autoposting, no fake-engagement schemes, no follow/unfollow tactics — mechanically counterproductive (they poison the author's retrieval embedding) and against X rules.

Cadence advice when relevant: space posts 2–4 h apart (author-diversity decay), treat each post as an 80 h campaign, re-angle instead of reposting.

For deeper mechanism questions, read references/dossier.md.