What Is RedSkill? How Xiaohongshu Skills Reach Your Agent

The short version

RedSkill is Xiaohongshu's distribution channel for AI agent Skills. It is not a website you browse. The official documentation you can actually fetch — redskill.xiaohongshu.net/install.md — documents exactly one thing: a shell command that installs a CLI called redskill into ~/.local/bin, and then installs Skills by identifier from the command line.

Everything else in circulation comes from Chinese tech media published around 2026-05-24, which describes the same system as an invitation-based beta "Skill store for AI Agents" with no interface of its own. Where those two evidence lines disagree, we say so below rather than picking the more exciting version.

What the official installer actually does

The official document gives one install command, plus a "CLI only" variant:

According to reporting from 53AI and Jimo Studio, that script does three things: it drops the redskill binary into ~/.local/bin; it writes two default Skills named find-redskills and redskill-preference into the agent workspace; and those two Skills tell the agent to prefer the RedSkill search-and-install path when a user asks to find or install a skill. The reporters call this preference injection.

If redskill is not found after install, the official fix is a PATH entry, not a reinstall:

The officially documented main path for getting a Skill is install-by-identifier into the current workspace:

Media reports add a wider command set — redskill search, list, uninstall, upgrade and self-upgrade — that the official document does not mention at all. Treat the extra commands as reported, not guaranteed: the same reporting found that self-upgrade was returning an empty bundle URL because the backend was not configured.

Two ways a person finds a Skill

1. Keyword search from the terminalThe reported flow is redskill search <keyword>, which returns an identifier, followed by redskill install <identifier>. The same reporting notes there was no endpoint that lists the whole store, so search results depend on the keyword you happen to type.
2. A passphrase pasted from a noteCreators can attach an approved Skill to a note. Readers copy a short natural-language instruction and paste it into their agent, which installs the CLI if needed and then the named Skill. A verbatim example from the reporting reads: "check whether RedSkill is installed; if not, install only the RedSkill CLI from the documentation, then install the space-weread-analyzer skill."

These are different discovery models with different consequences. Terminal search is a keyword gamble. The note path is editorially curated by whoever wrote the note — which means the recommendation carries the note author's incentives, not a ranking you can audit.

What a Skill physically is

The install artifact is a zip bundle. Reported behaviour: the CLI fetches the bundle address, verifies a sha256 hash, and unzips it into skills/<identifier>. The same reporting observed installs failing outright when the sha256 check was missing — which is worth knowing, because an integration that silently skips verification is a different risk class from one that refuses.

Inside the folder, the agent reads a SKILL.md instruction file written in natural language: trigger conditions, execution steps and dependency requirements. Two of the three media sources write the filename as SKILL.md; one writes skill.md. We use the uppercase form because more sources use it, and we flag the inconsistency rather than smoothing it over.

How a Skill gets published

The supply side runs through the Xiaohongshu creator console, and it is web-only — the app is not supported for upload. Reported steps: open the creator console, choose the Red Skill entry, upload a single Markdown file or a folder containing Markdown, let the system parse the name, description and body, fill in the identifier by hand, then submit for approval. After approval, a note can carry the Skill as an attachable component.

Two constraints from that reporting are important enough to repeat plainly. Only Markdown is parsed today, and script files are filtered out — if your Skill depends on Python or Node, the reporting says you have to wait. And upload itself was in a grayscale rollout while the copy-passphrase feature was open to everyone.

What we could not verify

This section is the reason to trust the rest of the page. These are things widely asserted about RedSkill that no source we could fetch actually supports:

• Store categories. No source lists a category taxonomy. Our own eight clusters on this site are our own grouping of a public ClawHub snapshot — they are research scaffolding, not platform categories.
• MCP. No source connects RedSkill to MCP. MCP-based Xiaohongshu tooling exists in the wider ecosystem; it is not part of the documented RedSkill distribution path.
• An official web store URL. None is documented. Fetching the bare official domain root returned HTTP 500 when we checked.
• Upload size and count limits, and pricing. Nothing stated anywhere we could reach.
• Windows support. One source says native Windows PowerShell is a poor fit and Linux, macOS or WSL are better; no other source mentions platforms at all.

Sources

1. Official install document — redskill.xiaohongshu.net/install.md. Installer URL, CLI-only flag, install-by-identifier as the main path, PATH fix.
2. 53AI, 2026-05-24 — xiaohongshu Red Skill beta report. Installer behaviour, default Skill names, command list, zip and sha256 flow, backend interface names, platform fit notes.
3. Jimo Studio, 2026-05-24 — Red Skill beta write-up. Corroborates the installer breakdown and the absence of a store UI. Shares wording with the 53AI piece, so we treat both as one evidentiary line.
4. xmsumi — Skill attachment report. Creator console upload steps, Markdown-only parsing, grayscale rollout, the passphrase flow into Claude Code and Cursor.

Where to go next