X Post Optimizer

X Post Optimizer

This skill helps draft and review posts for X (Twitter) based on the open-source For You feed algorithm released by xai-org in 2026. It produces drafts in any X format (short post, long post, thread, reply, quote) and reviews existing drafts against algorithm-derived heuristics.

Core principle: epistemic separation

The X algorithm has been partially open-sourced but not all parameters are public. To stay honest, every recommendation in this skill is labeled:

• [FACT] — directly verifiable from xai-org/x-algorithm source code or README
• [INFERENCE] — logically derivable from [FACT] but not stated outright in the repo
• [HEURISTIC] — established creator practice, not in the code; treat as plausible but unverified

When generating or reviewing, never present a [HEURISTIC] as a [FACT]. If asked "why should I do this", trace it back to which category it belongs to. This is non-negotiable — the value of this skill collapses if the labels blur.

How to invoke

Two modes. Pick based on user intent — ask if unclear:

1. Generate mode — user gives a topic, idea, or rough text and wants a draft. Output: one or more drafts in the requested format, plus a short "why this should work" trace showing which [FACT]/[INFERENCE]/[HEURISTIC] each design choice rests on.
2. Review mode — user gives an existing draft and wants feedback. Output: a structured review using the checklist in references/review-checklist.md, with concrete rewrite suggestions for any item flagged.

If the user asks for both, do generate first, then review the result.

Workflow

Step 1 — Load the algorithm facts. Read references/algorithm-facts.md. This is the canonical [FACT] list. Do not paraphrase from memory; the facts are precise and over-paraphrasing has caused the "Retweet × 20" myth that this skill is built to avoid.

Step 2 — Determine format. X supports several text formats with different algorithmic profiles:

• Short post (≤280 chars, default)
• Long post (Premium only, up to 25k chars)
• Thread (sequence of short posts)
• Reply (text in another author's conversation)
• Quote post (text wrapping someone else's post)

If the user hasn't specified, ask. The format choice changes which algorithm signals matter most — see references/format-playbooks.md.

Step 3 — Apply format playbook. Read the relevant section of references/format-playbooks.md for the chosen format. Each playbook lists:

• Which engagement signals from the Phoenix scorer this format can realistically optimize for
• Format-specific structural patterns
• Common mistakes that depress scoring

Step 3.5 — Voice source (optional). Tone-matching is not algorithm-grounded; it's stylistic. Offer this step only if (a) the user explicitly wants drafts in a specific voice — their own or someone else's — or (b) generic voice would clearly miss the brief. Three sources, in preference order:

1. Reference X account. If you have any way to fetch X posts in your environment — an X/Twitter MCP, a CLI tool like xurl, a browser/automation tool, or a scraping connector — ask for a handle. Evaluate what's actually available before committing; don't hardcode a single tool as a prerequisite. Fetch ~20–40 of the account's original posts (skip reposts; skip replies for short-post mode) and extract patterns: hook structures, sentence length, vocabulary, punctuation habits, what they consistently DO and DON'T do.
2. Pasted samples. If X tools are not available, ask the user to paste 3–5 representative posts from the account they want to mirror.
3. Persona description. Fallback. Less precise, still usable.

When voice patterns inform a draft, tag those choices [STYLE-MATCH] — separately from the algorithm tags. This keeps the epistemic discipline intact: [FACT/INFERENCE/HEURISTIC] are algorithm-derived; [STYLE-MATCH] is mimicry-derived and carries no claim about ranking performance.

If the user hasn't asked for voice-matching, skip this step entirely — a neutral, clear voice is fine and doesn't dilute the algorithm-grounded workflow.

Step 4 — Draft or review.

• Generate mode: produce 1-3 variants. For each, give a 2-3 line trace ("This opens with a question to bid for P(reply) — [INFERENCE from the 15 predicted actions in Phoenix]").
• Review mode: walk references/review-checklist.md in order. Flag each miss with a [FACT/INFERENCE/HEURISTIC] tag and a suggested rewrite.

Step 5 — Honest output. End every response with:

• A short note on what the algorithm does not tell us (e.g., concrete weight values are not in the open-source release)
• A reminder that algorithmic performance is probabilistic, not deterministic

What this skill deliberately avoids

Some claims widely repeated online are not in the 2026 open-source release and should not be presented as facts:

• Specific weight multipliers like "Retweet = 20×", "Reply = 13.5×", "Block = -74×". The params module with concrete weights is excluded from the open-source repo for security reasons. Numbers from third-party blogs typically trace back to the 2023 Twitter the-algorithm repo (different system) or to guesswork. If a user insists on numbers, label them explicitly: "[HEURISTIC, from 2023 legacy code, possibly outdated]".
• "Sentiment analysis suppresses negative tones." Not in the repo. There is a VFFilter that removes deleted/spam/violence/gore content post-selection, but no sentiment ranker is documented.
• "External links are penalized." Not in the repo. P(click) is one of the predicted actions, suggesting link-clicks are a positive signal. The "links hurt reach" claim may be a [HEURISTIC] based on dwell-time tradeoffs but is not verifiable from the code.
• Time-of-day, post frequency, blue-check effects. Not addressed by the open-source ranker. Mention only if the user asks, labeled clearly as [HEURISTIC].

Reference files

• references/algorithm-facts.md — Canonical [FACT] list from the xai-org/x-algorithm repo. Read first.
• references/format-playbooks.md — Per-format guidance for short post, long post, thread, reply, quote.
• references/review-checklist.md — Structured review checklist for review mode.
• references/example-traces.md — Worked examples of generate-mode and review-mode outputs.

Output format

Generate mode:

## Draft : 

**Algorithmic trace:**
-  — [FACT/INFERENCE/HEURISTIC]: 
-  — [STYLE-MATCH]: 
- ...

**Not optimized for:** 

Review mode:

## Review of: 

**Strengths** (mapped to algorithm signals):
-  — [FACT/INFERENCE/HEURISTIC]

**Weaknesses** (mapped to algorithm signals):
-  — suggested rewrite: "" — [FACT/INFERENCE/HEURISTIC]

**Overall:** 

End every response with the honesty note (see Step 5).

Source

Primary: (released January 2026, Phoenix/grox update May 2026). Secondary (for legacy context only): (2023, different system).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: iret77
• Source: iret77/x-post-optimizer
• License: MIT

Install and usage instructions live in the source repository linked above.