xiaohongshu-ai-workbench: Codex Skills for Xiaohongshu Operations

What xiaohongshu-ai-workbench actually solves

The repository is written for a narrow reader: someone with a product, service, course or real experience who cannot explain it clearly on Xiaohongshu. The README states the project is a companion to the paid handbook 《小红书运营手册》, and draws the line explicitly: the handbook handles judgement, the workbench handles execution. That division explains the shape of the codebase. There is no model, no crawler and no analytics dashboard here. There are seven skill folders, a packaging script and a validator.

The problem it addresses is that Xiaohongshu work is usually done from inspiration in one sitting: write a title, rewrite the bio, answer comments, then wonder why nothing connects. The README describes the skills as coming from 好事引力's repeated diagnostic, content planning and brand expression work, and argues that titles, profiles, topic selection, comment sections and conversion paths should be broken into actions that can be checked, reused and handed to an AI. That is a workflow claim, not a growth claim, and the README is blunt that it does not promise follower growth, viral notes, sales or platform algorithm results.

It is not for people who want a guaranteed result, who want to paste in someone else's case study or a celebrity quote, or who have no product material and want the AI to invent an account from scratch. The README lists all three under who it is not for.

Seven skills, one workflow each, routed by a parent skill

Each skill is a self-contained folder at the repository root: xiaohongshu-suite, xiaohongshu-title, xiaohongshu-profile, xiaohongshu-magazine, xiaohongshu-topic-planner, xiaohongshu-comment-reply and xiaohongshu-conversion-path. The README's table gives each one a single job. xiaohongshu-title generates, diagnoses and optimises titles. xiaohongshu-profile audits the profile and rewrites the bio. xiaohongshu-topic-planner builds topic pools and series. xiaohongshu-comment-reply handles replies and pinned comments. xiaohongshu-conversion-path connects content, profile, direct messages and product. xiaohongshu-magazine, added in v0.2.0, treats the account as a magazine: it sets a 刊魂 (the issue's core theme), defines columns, and produces a reusable topic library in that style.

xiaohongshu-suite is the router. The README's example prompt asks it which of profile, topics or conversion path to fix first, and the recommended workflows are written as arrow chains: profile to topic-planner to title for a cold start; topic-planner to title to comment-reply for a single new note; profile to conversion-path to topic-planner to title when selling a service. The routing is plain text, so the mechanism is prompt composition rather than code execution. The validator confirms this architecture is enforced: it checks that every subdirectory has a valid SKILL.md, that the skill name matches the directory name, and that agents/openai.yaml metadata exists. A skill that solves two workflows would fail the project's own contribution rules.

Installing the .skill files and running a first profile check

There are two paths. The README says that if you only want to use the skills, you download the .skill files from packages/ and import them into Codex. The listed packages are xiaohongshu-suite.skill, xiaohongshu-title.skill, xiaohongshu-profile.skill, xiaohongshu-magazine.skill, xiaohongshu-topic-planner.skill, xiaohongshu-comment-reply.skill and xiaohongshu-conversion-path.skill. The second path is for people who want to modify the prompts: clone the repository and run the packaging and validation scripts from the repository root.

python3 scripts/package_all.py
python3 scripts/validate_all.py

The first command writes packaged files into packages/. The second checks the same things the contribution rules describe: valid SKILL.md in every subdirectory, matching names, agents/openai.yaml present, packaged files consistent with source, no stale external-source references in README, skill bodies, OpenAI metadata or eval samples, and the project's brand marker present in README, LAUNCH, every SKILL.md and the OpenAI metadata. If the validator fails, the packaging step is the thing to re-run after fixing the offending file.

The README also documents a manual route. You copy the skill folder you need into ~/.codex/skills/, so that a single skill lands at a path like this:

~/.codex/skills/xiaohongshu-title/SKILL.md

After installation, the README's profile example is the most direct first use. It expects you to fill in the fields rather than describe the account in prose: 昵称, 简介, 置顶笔记, 我卖的产品/服务, 目标用户. The output you should expect is a profile audit plus a rewritten bio, not a list of growth tactics. The README does not document rollback or an uninstall step, and it does not describe how Codex reports a skill that fails to load.

No case library, no formulas, and that is deliberate

The README states that the project does not ship external cases, an external formula library or third-party endorsements, and that contributing rules forbid adding them. For an operator used to prompt packs stuffed with screenshots of other people's viral notes, this is a real constraint: the skills have nothing to imitate. The input quality is the output quality. If you cannot describe your target user, your product and your delivery, xiaohongshu-topic-planner has nothing to plan from, and the README places that situation under who it is not for.

The second limitation is scope. The README says the skills fit Xiaohongshu and RED social scenarios and are not generalised into an ordinary writing tool. That means the prompts carry platform-specific assumptions about profiles, pinned notes, comment sections and direct-message conversion. Using xiaohongshu-title to write a newsletter subject line is outside what the project claims to do, and the validator's brand-marker check suggests the maintainers intend to keep it that way.

The third is maintenance. The last push to the default branch was on 2026-08-07, and the most recent release, v0.2.0, added xiaohongshu-magazine on 2026-08-04. The README's contribution rules require a new skill to come with agents/openai.yaml metadata and at least three eval samples, which sets a floor for anyone extending it, but the README does not describe a release cadence or a deprecation policy for existing skills.

How this differs from a general prompt library

The closest alternative is a generic prompt collection for content creation, the kind distributed as a Notion page or a folder of text files. The difference is structural. A prompt library gives you a block of text to paste and edit; xiaohongshu-ai-workbench gives you installable Codex Skills with a fixed folder layout, a name that must match the directory, agents/openai.yaml metadata and eval samples. That layout is what lets xiaohongshu-suite route to the other skills by name and what lets scripts/validate_all.py check that packaged files match source.

The second difference is the refusal to bundle examples. A generic library usually ships filled-in samples so the output looks plausible immediately. Here the README's real-use examples are all input templates: a freelancer asking for a seven-day topic plan for a quotation calculator, a service IP asking for a profile check for brand expression diagnostics, a free mini-program asking for a path from content to trial. The user supplies the substance. That is slower at the start and harder to fake, which is consistent with the README's statement that the project is not a substitute for the handbook.

A third difference is the separation of diagnosis from output. The design principles say each skill should diagnose the problem first and then give an executable output. A prompt library typically goes straight to generation.

Licence, packaging and the cost of staying current

The repository is MIT licensed. The README states you may use, modify and distribute it freely as long as the licence is retained. For an operator, that means the prompts can be edited in place, folded into an internal workflow or redistributed to a client, provided the LICENSE file travels with the copy. It does not mean the companion handbook is free: the README links the handbook at xiaobot.net/p/xiaohongshuku as a separate paid product, and the licence covers the code and prompts, not the book's content. Nothing here is legal advice; read LICENSE and CONTRIBUTING.md before redistributing a modified package.

Upgrade cost is low by design. Skills are folders, and packages/ holds the distributable .skill files, so upgrading means replacing a folder or re-importing a package rather than migrating a database. The maintenance work sits on the contribution side: adding a skill requires agents/openai.yaml and at least three eval samples, and scripts/validate_all.py will fail the build if the brand marker is missing from README, LAUNCH, any SKILL.md or the OpenAI metadata. That validator is the project's main defence against prompt drift, because it also checks that no stale external-source references have crept into skill bodies or eval samples.

Editorial conclusion

Adopt it if you already have a real product or service and want Xiaohongshu profile, title, topic, comment and conversion work broken into separate prompts you can run one at a time. Skip it if you want guaranteed follower growth, ready-made case studies or an account invented from nothing; the README rules those out, and the skills ship without external case libraries. Verify first that your Codex setup loads a .skill file from packages/ or a folder copied into ~/.codex/skills/, and read xiaohongshu-title/SKILL.md before trusting its output on a live account.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. mengke-wang/xiaohongshu-ai-workbench on GitHub
4. README
5. Releases