AgentSkillsCN

x-safety-filtering

在分析内容为何未被展示、排查是否存在“影子封禁”情况，或理解平台的安全合规层时，可运用此技能。该技能涵盖了 `VisibilityLib` 逻辑——它在候选内容筛选之后、最终排名之前运行，旨在保障平台的健康运行。

SKILL.md
--- frontmatter
name: x-safety-filtering
description: Use this skill when analyzing why content is not appearing, checking for "shadowban" conditions, or understanding the safety compliance layer. This skill covers the `VisibilityLib` logic that runs *after* candidate sourcing but *before* final ranking to ensure platform health.
version: 1.0.0
license: MIT

X Safety & Filtering

Expert knowledge of X's VisibilityLib, covering safety labels, shadowban logic, NSFW filtering, and content health suppression mechanisms.

Context

The filtering stage is the "Gatekeeper" of the timeline. Even if a tweet has a high ML score from the Heavy Ranker, VisibilityLib can drop it entirely or apply a "Do Not Amplify" label that restricts it to the author's profile. This layer enforces legal compliance, user blocks, mutes, and platform safety rules (e.g., NSFW or Toxicity).

What it does

  • •Decodes "Shadowbans": Explains the specific internal labels (like SearchBlacklist) that cause users to perceive they are shadowbanned.
  • •Enforces User Preferences: Handles the logic for Mutes, Blocks, and "Show less often" signals.
  • •Manages Content Health: Identifies toxic content or misinformation using models like pToxicity and pAbuse and applies downstream penalties.
  • •NSFW Handling: Segments content into "Adult" or "Graphic" categories using pNSFWMedia and ensures it respects the viewer's sensitivity settings.

Guidelines

  • •SafetyLevel Context: Rules are evaluated based on the SafetyLevel (e.g., Timeline vs. Profile). A tweet might be visible on a Profile but blocked in the Home Timeline.
  • •The "Do Not Amplify" (DNA) Label: Disqualifies tweets from the "For You" (Out-of-Network) timeline and Search results without removing them from the profile.
  • •Visibility vs. Ranking: 1. Pre-Scoring: Hard filters (Drop) remove blocked or legally prohibited content. 2. Post-Scoring: Soft filters (Labels) apply safety checks (e.g., author diversity) after the Heavy Ranker has assigned scores.
  • •Toxicity Thresholds: If a user enters a "Reply Guy" mode with consistently high pToxicity scores, their account enters a state that limits the reach of all their future replies.
  • •Linear Decay: Negative reputation signals follow a linear decay model; an account can "heal" its reputation over time by stopping negative behavior.

Example Trigger Prompts

  • •"/safety-check shadowban or SearchBlacklist status"
  • •"/safety-check toxicity decay for @user"
  • •"/safety-check flagged content vs safe content ratio"
  • •"/safety-check audit applied filters on a thread"
  • •"what are the visibility rules affecting recent posts"
  • •"show suppressed accounts in a community"