Creator Buddy earns credibility by documenting its own failure modes

Creator Buddy is a collection of agent skills for Chinese-platform content creators — specifically WeChat Official Accounts (公众号), Xiaohongshu, and video — built on the open Agent Skills protocol. The design philosophy is worth stating plainly: it refuses to be a "one-click account launcher." Instead it splits every workflow into two phases — intelligence gathering first, then production — and insists that the content remains yours throughout.

The intelligence half pulls real platform data before any writing begins. A search for trending Codex content on Xiaohongshu, for instance, returns scored, deduplicated results with engagement counts, then surfaces a user-language insight: the audience there cares about installation cost and accessibility, not model benchmarks. That grounding step is what separates the tool from a prompt that just asks an LLM to invent ten topic ideas from nothing.

The production half is genuinely broad. For WeChat alone there are eleven distinct skills covering account positioning, six different long-form writing modes, short posts with explicit anti-AI-voice style rules, title generation with A/B scoring, three separate image/diagram skills, and full HTML layout output. Xiaohongshu gets ten skills with a master controller that diagnoses which stage of account growth you're stuck at and routes accordingly. Video gets a full pipeline from topic research through B-roll generation — including a hand-drawn animation path via Seedance — subtitle transcription from YouTube, Bilibili, Douyin, and Xiaohongshu links, voiceover mixing, and cover design.

The honesty section at the bottom is unusually direct for a README. It explicitly states that trending content only tells you what spreads on a platform, not what fits your positioning; that generative images and video can produce visible errors requiring human review; that Douyin support is currently just an interface stub awaiting a local read-only CLI; and that Xiaohongshu detail pages depend on the full URL including the xsec_token returned by search. A tool that lists its own failure modes in the documentation is worth taking seriously.

Security posture is also spelled out: read-only public data only, no posting or account actions, credentials stay in local environment variables and never enter the repository, and low-frequency sampling is explicitly recommended to avoid triggering platform rate limits.

Installation is a single npx skills add command that auto-detects the running agent runtime — Claude Code, Codex, Cursor, and several others are listed as compatible. Each skill is a self-contained directory with a SKILL.md and optional scripts, so manual copy-paste into any conversation works as a fallback even without runtime support.

The scope is ambitious and the architecture is coherent. Whether the external data dependencies — Redfox, Agent Reach, Guaikei API — hold up in practice is something the README cannot answer.