Helps individual content creators fight against Claude's common "English-style thinking" contamination and produce Chinese content that truly has emotion and internet feel.
Core Problem (PAIN-301 empirical evidence)
In complex content generation tasks, Claude easily falls back into English chains of thought, leading to:
- Stiff sentence patterns ("Dear friends, can you understand this feeling?")
- Westernized cultural frameworks (American TV show memes, In conclusion, Western linear logic)
- Diluted emotion
- The probability of failure midway through long content rises sharply
Community consensus: Claude is good at making frameworks,internet feel needs extra protection。
Trigger Scenarios (Real Users' Exact Words)
When the user says any of the following, activate this skill immediately:
- "Claude's Xiaohongshu writing has too much translationese" "English-style thinking output is a disaster"
- "Help me rewrite this note into pure Chinese internet feel" "Remove Claude's translationese"
- "The WeChat official account article turned into English midway" "50 Skills content generation failed"
- "Force Claude to think in Chinese when writing notes" "Polish emotional expression for Xiaohongshu"
- "As it writes, it turns into a foreigner pretending to write in Chinese"
- "The first 800 characters are okay, then the next 800 characters go straight to In conclusion"
Complete Workflow (Forced Chinese Chain of Thought + Detection + Rewriting)
Phase One: Force a Pure Chinese Chain of Thought (The Most Critical Defense)
Before starting any planning, structure, or examples, you must first output a passage ofcomplete Chinese internal monologue：
- "I am now a senior Xiaohongshu/WeChat official account blogger with 100,000+ followers, using my Chinese brain to recall my real personal experiences, awkward moments, and details readers will resonate with."
- Write in Chinese: What is the real moment in this topic that moves me most? What is the reader's most painful point? What is the emotional curve? What catchphrases would be used to express it?
Strictly prohibitedto use English thinking and then translate at this step.
Phase Two: Immediately Conduct English Leakage Detection After Generating the First Draft
Scan paragraph by paragraph for the following translationese markers (hard indicators + cultural frameworks):
- Hard translationese sentence patterns: "Dear friends, can you understand this feeling?" The stiff version of "Sisters/brothers, who gets it" "This thing is just too XX!"
- Western frameworks: "In conclusion / Furthermore / It is important to note / We should consider"
- Cultural mismatch: American TV show memes, Hollywood narrative, individualist hero stories, linear "problem-solution-conclusion"
- Rhythm problems: almost all sentences are the same length, lacking real first-person details of "failure/heart flutter/regret"
Detection results must output:Leaked fragment annotations + severity level。
Phase Three: Internet-Feel Rewriting (Inject Real Corpus)
For each problematic paragraph detected:
- Force a rewrite with a Chinese-thinking brain: "Recall your own real moments as a blogger, and rewrite this passage with colloquial speech + emotion + short sentences + a sense of pauses"
- Inject
references/真实高赞网感示例.mdthe real high-engagement expressions ("绝绝子", "泪目", "yyds", "姐妹们冲就完事儿了") - Adapt per platform (see
references/平台网感适配器.mdfor detailed rules)
Stage Four: Multiple Versions + Final Self-Check
Output 2-3 versions with different emotional intensity (conservative / standard / maxed-out emotion). Finally, run the checklist one more time and output:
- Internet-savvy score (1-10)
- Remaining leaked fragments
- Change comparison table (Before → After + why this change makes it feel more human)
Stage Five (optional but recommended): Generate a persistent guard fragment
If the user uses it frequently, automatically generate an "anti-English-thinking system prompt" and suggest the user paste it at the very top of CLAUDE.md.
Recommended "Chinese Internet-Savvy Guardian Protocol v1.0"
Fully copyable protocol see references/中文网感守护协议.md。
Strongly recommended: all writing-type Skills (including ones you write yourself) should, in the "write copy" stepbeforereference this protocol.
For platform-specific tone differences and adaptation Prompt templates, see references/平台网感适配器.md。
Strongest combined workflow (most effective in practice)
- First call the
xiaohongshuskill → get the framework + platform hard constraints + 7-3 principle + real data - Feed xiaohongshu's output tothis skillto do "forced Chinese CoT + full cleanup of translationese + real internet-savvy injection"
- Optionally then use
humanizerfor a final layer of de-AI-flavor polishing (24-pattern self-audit)
This chain orthogonally combines "Claude's framework capability" + "Chinese internet savvy" + "de-AI-flavor", and in practice works far better than using any one of them alone.
Safety Contract (never violate)
- Only change language and cultural expression, never change the core viewpoints, structure, or data the user originally intended to express
- Always respect the persona, tone, and forbidden words the user has set in CLAUDE.md / AGENTS.md
- Providemultiple versionsfor the user to choose from, rather than forcibly replacing the original draft
- All modifications must providea traceable change comparison + reasons for the changes
- When the user's original draft is found to already have very strong internet savvy, prioritize outputting "your original draft already has great internet savvy, here are some fine-tuning suggestions", and never over-polish
Health and Version Statement
Current version：v1.0 Last manual verification: 2026-05-28 (Claude Code 20250514 + Spellbook main branch) Known limitations: This skill does not conjure internet savvy out of thin air; it relies on the real highly-liked corpus anchors provided by the user + the user's own persona settings. What it is best at ispreventing Claude from translating content that originally had internet savvy into content that has none。
It is recommended that whenever a new case of translationese is found, immediately add it to references/真实高赞网感示例.md and the checklist.
Limitations statement (honestly inform the user)
- Limited effect on extremely short content (<150 characters) (not enough context to judge internet savvy)
- Cannot replace the user's own real life experiences (this is the fundamental ceiling of all AI writing)
- If the user themselves already likes the "translationese academic style," this skill will proactively remind them: "It has been detected that you may prefer this style. Do you still want to force-convert it to internet-native feel?"
Evaluation philosophy (subjective-type Skills must comply)
Because "internet-native feel" is extremely subjective, this skillrefusesto use hard assertions as the sole evaluation criterion.
Recommended evaluation path (fully following skill-creator guidance):
- 70% weight: manual blind testing (using eval-viewer + blind comparator, having real content creators or independent Agents blindly choose "which piece reads more like it was written by a real person")
- 30% weight: lightweight proxy (English leakage rate regex scan + platform structure compliance count + first-person authentic detail occurrence rate)
- Gold standard: interaction data 7 days after actually posting to Xiaohongshu/WeChat Official Account (likes, saves, comments, completion rate) compared A/B against historical similar posts
Strictly prohibited overfitting to the test set in order to make evals pass. Refer to the PAIN-401 lesson: a skill that is perfect in evals but terrible for real users is a grave.
The correct way to use this skill: The user pastes a piece of content that "turns into translationese as they write" → this skill outputs a "detection report + 3 rewritten versions + why it was changed this way."
The user will immediately feel: "Finally, someone has pierced through this layer of window paper."