Huo15 Xiaohongshu

有记忆、能学习、会教方法的小红书创作助手。两套打分叠加 — ①工程师流（标题/首段/排版/emoji/话题/合规）②Allen 流（留白/AI腔/带读者/共鸣/邀请语/范本范，含 Jarvis 陷阱 5 维），加风格档案、规则覆盖、写作教练（一次只 focus 一维的渐进式 / 全维诊断 / LLM 改写）、对话...

Security Scan

The skill mostly matches its stated Xiaohongshu (小红书) authoring purpose, but there are inconsistencies around credentials, metadata, and a prompt-injection signal that merit caution before use.

Purpose & Capability

The name/description align with the included scripts (scrape, analyze, write, coach, compliance, publish-helper, A/B tests). However the package metadata declares no required environment variables while the README/SKILL.md explicitly tells users to export an XHS_COOKIE and optionally LLM keys (e.g., ANTHROPIC_API_KEY). That mismatch (declared none vs. actual need for cookies/API keys) is an incoherence that should be resolved by the publisher.

Instruction Scope

Runtime instructions and README tell the user to provide a full Xiaohongshu cookie and to run scraping scripts (scrape-note/search/user) and to persist data under ~/.xiaohongshu. Those actions are within the stated research/authoring scope, but they require sensitive credentials and write persistent local files. Additionally a pre-scan flagged unicode-control-chars in SKILL.md (possible prompt-injection artifact) — that is unrelated to the claimed feature set and raises concern about embedded instructions that could manipulate LLM behavior.

Install Mechanism

No install spec is provided (instruction-only install), and all code is bundled in the skill. There are no network download/install steps in the registry metadata, which reduces installer risk. The code will run locally when invoked; no external arbitrary installer URLs were observed.

Credentials

The skill requires (per README & SKILL.md) the user's Xiaohongshu cookie (XHS_COOKIE) to fetch protected pages and optionally an LLM API key (e.g., ANTHROPIC_API_KEY) for LLM-powered features. Those secrets are proportionate to the scraping/LLM augmentation use cases, but the registry metadata does not declare them. The skill also persists tokens/note IDs and may accept xsec tokens in A/B test commands — all of which deserve explicit declaration and careful handling by the user.

Persistence & Privilege

The skill stores user data under ~/.xiaohongshu/ (drafts, profile, snapshots, ab_tests.jsonl). That persistence is expected for a creator tool. It does not request always:true or system-wide modification privileges and explicitly states it will not automate publishing; it modifies only its own data area.

Scan Findings in Context

[unicode-control-chars] unexpected: A prompt-injection pattern was detected in SKILL.md. There is no legitimate need for hidden/unprintable control characters in runtime documentation; this could be an artifact of packaging or an attempt to influence LLM behavior. Inspect SKILL.md raw content and remove/verify any control/unprintable characters before trusting the skill.

What to consider before installing

What to check before installing/running this skill: - Secrets: The README and SKILL.md instruct you to export XHS_COOKIE (your Xiaohongshu session cookie) and optionally an LLM API key. Those are sensitive; only provide them if you trust the source. The registry metadata failing to declare required env vars is an inconsistency — ask the publisher why it's omitted. - Review code locally: This skill contains many scraping scripts (scrape-search/note/user, xhs_client.py) that will make requests using your cookie and will persist data to ~/.xiaohongshu/. Open and review xhs_client.py, scrape scripts, and any networking code to confirm traffic goes only to xiaohongshu.com or to known LLM providers (anthropic) and that there are no hidden external endpoints or URL shorteners. - Prompt-injection artefact: The SKILL.md had unicode-control-chars detected. Inspect SKILL.md and other text files for unprintable characters or suspicious injected prompts that could try to influence an LLM-runner's behavior. - Limit exposure: If you want to try the tool, run it in an isolated environment or on a throwaway account; do not reuse your primary account cookie if you are unsure. Consider using a browser session with minimal privileges or tooling that can issue read-only API tokens (if available). - LLM keys: If you provide an Anthropic (or other) API key, be aware your drafts or scraped content may be sent to that provider. Confirm what data is sent and whether prompt caching/telemetry behaviors are acceptable. - Auto-publish assurance: The skill claims "绝不自动化发布" — still verify publish_helper.py to ensure no code will auto-post or leak credentials. - Ask the publisher for fixes: Request that the package metadata explicitly list required env vars (XHS_COOKIE, optional XHS_LLM_PROVIDER & ANTHROPIC_API_KEY) and an explanation for the prompt-injection flag. If the publisher cannot justify these, treat the skill as higher risk. If you cannot validate these points, avoid running the scraping/publishing parts and consider using only the offline analysis/writing helpers after manual inspection.

火一五小红书创作伙伴 v3.2

详细文档见 README.md，版本历史见 docs/changelog.md。

能做什么

阶段	命令
入门	assistant.py init --baseline ... 建风格档案
状态	assistant.py status / next / today
调研	safety_check.py → scrape-{search,note,user}.py → analyze-notes.py
选题	topic_ideas.py / brainstorm.py / today.py
对标	reverse_engineer.py --url <爆款>（拆出公式/骨架/Allen 6 维/why it works）
创作	write_post.py draft 出骨架 → drafts.py 版本管理
教练	coach_iterate.py 一次 focus 一维 ｜ coach.py 全维 ｜ critique.py Allen 美学 ｜ polish_post.py 工程分
改写	critique.py --rewrite（需 LLM）｜ practice.py rewrite-jarvis 训练改写思路
配套	coin_word.py 造词 ｜ series_design.py 栏目化 ｜ reader_simulate.py 多读者 ｜ cover_brief.py 封面
合规	compliance_check.py（绝对化词/医美/导流/诱导互动）
发布	publish_helper.py（剪贴板 + 10 项 checklist；不自动化）
复盘	track_post.py snapshot → weekly_review.py
训练	`practice.py prompt
学习	assistant.py learn key=value ｜ evolve ｜ `preset allen

不做什么

❌ 自动发布 / 多账号矩阵 / AI 生图 / 一键全文 / 达人投放分析

防封号红线

1. 用自己的 Cookie，脚本不做登录自动化
2. 每次请求 37 秒延时，单会话 30 次封顶，会话间隔 1030 分钟
3. 460 / 461 / 403 / captcha / 重定向登录 → 立即停 30 分钟
4. 不翻页批量抓（搜索只取首页，主页只取 preview）
5. 日请求 ≤ 100 次

工作流（全部走 assistant.py）

# 第一次
python3 scripts/assistant.py init --persona "30+ 干皮女生" --niche "护肤" \
    --baseline note1.md note2.json
python3 scripts/assistant.py preset allen        # 切 Allen 美学加权

# 每天写一条
python3 scripts/assistant.py today               # 今日选题
python3 scripts/assistant.py drafts new --topic <主题>
python3 scripts/assistant.py write <主题>        # 起草
python3 scripts/assistant.py coach-iterate <id>  # 一次只改一维（推荐）
python3 scripts/assistant.py drafts diff <id>    # 对比 v_n vs v_{n-1}
python3 scripts/assistant.py publish <id>        # 剪贴板 + checklist

# 看到爆款想学
python3 scripts/assistant.py reverse --url <URL> --add-baseline

# 周末
python3 scripts/assistant.py review              # 周/月复盘
python3 scripts/assistant.py learn disable=emoji # 教助手新规则
python3 scripts/assistant.py evolve              # 自动演进

Allen 文案心法

1. 「好文案不是写出来的，是留出来的。」— 留白
2. 「站文案里面读文案，不是站在外面分析。」— 第三课
3. 「卖的是身份认同，不是商品本身。」— 第二课
4. ❌ 教读者「怎么做」 vs ✅ 展示「什么样的人已经在做」 — Jarvis 陷阱核心

详见 data/allen_method.md（Allen 三课 + 五技法 + 11 案例 + 关键认知转变 + Jarvis 陷阱 5 维差距）。

风格预设

预设	工程权重	Allen 权重	适合
allen	50%	50%	品牌号、情感号、生活号
engineer	100%	0%	干货号、教程、工具测评
balanced	70%	30%	综合个人号（默认）

python3 scripts/assistant.py preset allen

数据资产

data/ 目录包含 9 份对人和 Claude 都可读的资产：标题公式 11 种、正文骨架 7 种、 emoji 调色板、话题标签库、社区规则、敏感词、Allen 方法论、AI 腔黑名单、节气画面库。

个人档案位置

~/.xiaohongshu/
├── posts.jsonl / snapshots.jsonl  # 起草历史 + 互动快照
├── drafts/<id>/v01.md, v02.md, ... # 草稿版本（v3.2）
└── profile/
    ├── style.json / rules.json    # 风格 + 规则覆盖
    ├── baseline/ feedback.jsonl   # 代表作 + 反馈
    ├── iter_sessions/             # 渐进式教练历程（v3.2）
    └── reviews/                   # 周/月复盘归档

触发词

小红书 / xhs / xiaohongshu / 写小红书 / 小红书文案 / 爆款文案 / 小红书选题 / 小红书发布 / 小红书复盘 / 小红书教练 / Allen 流 / 范本范

---

重要免责： 仅用于合规、针对公开可见内容的个人调研与创作辅助。 请尊重 xiaohongshu.com 的服务条款。绝不支持商业批量采集 / 内容搬运 / 绕过风控 / 自动化互动。

技术支持： 青岛火一五信息科技有限公司