Dirtytrii Content Style Calibration Loop
Purpose
This skill turns user edits into reusable style rules. It prevents the agent from doing the shallow version of learning:
用户用了几个词 → 下次多塞这几个词
The correct goal is:
用户改了什么关系、视角、节奏、风险边界或人群钉子 → 下次按这个写作判断工作
Use this skill after the user edits generated text, rejects a tone, or says a draft does not sound like them.
When To Use
Use it when the user:
- edits a model-written paragraph and asks you to learn the style;
- says
不像我,太编辑,太 AI,太正经,太抽象,没网感,像 README, or像发布会; - distinguishes between a surface feature and the real rule, e.g.
不是多用这些词，是要有网感; - wants the lesson added to a skill, reference, README, or workflow;
- is calibrating a recurring content voice for Xiaohongshu, Douyin, WeChat, scripts, captions, or other public copy.
Do not use it to preserve raw private conversations. Extract rules, not transcripts.
Calibration Workflow
1. Capture the before/after pair
Get the model draft and the user-edited version. If the original is not available, still analyze the user edit, but mark the missing comparison.
Completion criterion: you can point to at least one changed phrase, deleted structure, or added line.
2. Diff meaning, not just wording
Classify each important change:
| Change type | Question |
|---|---|
| Voice | Did the user make it rougher, sharper, calmer, more local, more personal? |
| Viewpoint | Did the user move from expert explanation to participant experience? |
| Relationship | Did the user show what the tool/project does to a person? |
| Specificity | Did the user replace generic terms with an incident, object, role, or scene? |
| Rhythm | Did the user shorten, fragment, add breath, or keep a deliberate rough edge? |
| Risk boundary | Did the user remove over-cautious disclaimers or add factual anchors? |
| Platform fit | Did the user make it more suitable for XHS/Douyin/WeChat/etc.? |
Completion criterion: every accepted user edit is mapped to at least one reason.
3. Extract rules, not word lists
Bad extraction:
User likes 疑似、绷不住、抽象.
Better extraction:
User wants high-intensity internet-native judgment: identify the absurd point in a real situation, then say it with half-joke/half-accurate technical self-mockery. Internet words may appear, but should not be forced.
Completion criterion: the rule would still work if none of the original words were reused.
4. Keep the strongest examples
Store 1-3 compact before/after examples or pattern examples. Do not store entire chat logs, platform metrics, credentials, personal directories, or private context.
Completion criterion: each example demonstrates a rule and is safe to publish or share.
5. Decide the destination
Use the smallest durable destination that will affect future behavior:
| Lesson type | Destination |
|---|---|
| User preference that should follow all sessions | memory / user profile |
| Account-specific writing style | account style guide or content repo reference |
| Reusable agent procedure | skill |
| One-off task state | session notes or current repo files, not memory |
| Raw operational data | operations repo only, not public skills |
Completion criterion: the lesson is stored where the next relevant agent/tool can read it.
6. Verify the next prompt can use it
Write a tiny future-use prompt fragment, for example:
Use the calibrated style rule: websense comes from identifying the absurd point, not stuffing meme words. Keep technical anchors real, but open with the human friction.
Completion criterion: the prompt fragment is specific enough to change a future draft.
Output Format
When reporting a calibration result, use:
# 风格校准结论
## 1. 用户改动
## 2. 改动背后的规则
## 3. 下次写作要做
## 4. 下次写作不要做
## 5. 可沉淀位置
## 6. 已沉淀 / 待用户确认
Keep it short. The goal is not literary analysis; the goal is future behavior change.
Example
User changes:
我就是那个有无数想法、但被上班榨到一行代码都不想写的技术中登。
into:
轻松干翻我这种有无数想法、但被上班榨到一行代码都不想写的技术中登。
Do not extract:
Use the phrase 技术中登 more.
Extract:
Do not merely introduce the persona. Show the power relationship: the tool/AI/project has so much energy that it embarrasses, replaces, overwhelms, or exposes the person who was supposed to control it.
Public-Skill Hygiene
When turning a private calibration into a public skill:
- Keep general rules and safe examples.
- Remove platform backend data, engagement metrics, account IDs, session IDs, local paths, cron IDs, cookies, screenshots, and raw comments.
- Do not quote large private conversation chunks.
- Do not publish a user's personal biography unless they explicitly made it part of the public account voice.
- Prefer
user-edited sample/confirmed style samplewording over private chat provenance.
Common Pitfalls
- Learning the vocabulary instead of the judgment. If the user says
不是这些词，是网感, immediately rewrite the rule around judgment. - Overfitting one post. A high-performing post gives a hypothesis, not universal law. Record it as a style pattern to test.
- Saving too much. Skills should contain compact rules and examples, not raw drafts, raw chats, or platform data dumps.
- Skipping verification. A calibration is not complete until it produces a prompt fragment, checklist item, or skill/reference update that future agents can actually use.
- Letting the editor override the user. The agent may explain why an edit works, but the user's taste is the final style signal for their own account.