writing-style-skill

Writing Style Skill (Template)

Fork this skill and change it to your writing style. Built-in auto-learning, the more you use it, the more accurate it becomes.

🎯 How to use

Fork / clone this skill
Change the style rules below to your own
Have the AI write content using this skill
You edit until you're satisfied → the script automatically learns what you changed
Next time, what the AI writes will be more like you

[0] Voice Dimensions (Quantify Your Style)

Use a 1-10 score to define your style dimensions. The AI understands numbers more easily than phrases like "write a bit more naturally."

Dimension	Score	Your description
formal_casual	?/10	More formal or more casual?
technical_accessible	?/10	Technical depth?
serious_playful	?/10	Serious or lively?
concise_elaborate	?/10	Concise or detailed?
reserved_expressive	?/10	Conservative or direct?

💡 Don't know what to fill in? Run AI writing a few times first → you edit → auto-learning will help you fill it in.

[1] Role and Reader

Who I am:

(fill in your identity, e.g.: indie developer, newly minted AI enthusiast)

Who the reader is:

(fill in your reader, e.g.: technical people interested in AI)

Relationship with the reader:

(e.g.: peer exchange, not teaching)

[2] Writing Rules

Basic rules

(fill in your rules, e.g.: don't use "delve into", paragraphs should be short, include specific numbers)

Banned words

(words AI loves but you don't, e.g.: it's worth noting, in summary, this article will introduce)

Sentence structure preferences

(sentence structures you like, e.g.: conclusion first, no academic lead-in)

💡 You don't need to finish writing these from the start. These rules accumulate automatically through your edits. After 10 writing→editing cycles, dozens of precise rules will have grown here.

[3] Format Specifications

Platform adaptation

Platform	Format requirements
X/Twitter	Doesn't render markdown, use plain text
Xiaohongshu	Lots of emoji, short paragraphs
Blog	Standard markdown
(your platform)	(your requirements)

🔄 Auto-Learning (Built-in)

This skill automatically learns from your edits. No need to write rules manually.

How it works

AI 用这个 skill 写初稿
    ↓
你改到满意
    ↓
脚本 diff 两版 → 提取你改了什么
    ↓
新规则自动写入这个 SKILL.md
    ↓
下次 AI 写出来就更像你

Only two data points are needed

original: the first version generated by the AI
final: the version you finally confirm

It doesn't matter how many rounds of edits happened in between. Edited back and forth 10 times in Google Doc? Doesn't matter, only the first and last are compared.

Agent Operation Guide

After writing content:

python3 scripts/observe.py record-original <file> --account <账号> --content-type <类型>

After the human confirms the final version:

python3 scripts/observe.py record-final <file> --match <hash>

Extract rules (manually or automatically via cron):

python3 scripts/improve.py auto --skill .

Rule tiers

Tier	Meaning	Handling
P0	High confidence (appears multiple times)	Apply automatically
P1	Medium confidence	Manual confirmation
P2	Low confidence (only 1 time)	Archive and observe

Safe

Automatically back up SKILL.md before each update
improve.py rollback One-click rollback
auto mode only applies P0

📊 CLI Reference

observe.py (zero-dependency pure Python)

Command	Function
`record-original <file>`	Record AI draft
`record-final <file> --match <hash>`	Record final version
`pending`	View pending pairs
`stats`	Statistics

improve.py (requires LLM CLI)

Command	Function
`extract [--days 7]`	Extract improvement suggestions
`auto`	Extract + automatically apply P0
`show`	View proposals
`apply <id>`	Apply proposals
`rollback`	Rollback

Supported LLM CLI: claude（Claude Code）/ llm（pip install llm）/ IMPROVE_LLM_CMD Environment variables

📂 Data storage

~/clawd/memory/                    # OpenClaw
~/.claude/memory/                  # Claude Code
├── skill-runs/<skill-name>/
│   └── YYYY-MM-DD.jsonl          # 每日观察日志
├── skill-proposals/<skill-name>/
│   └── YYYYMMDD-HHMMSS.md       # 改进提案
└── skill-backups/<skill-name>/
    └── SKILL-YYYYMMDD-HHMMSS.md  # 自动备份

Automatically detects the environment; no manual path configuration required.

🚀 30-day expectations

Time	Expected effect
Week 1	Accumulate 3-5 edits and generate the first batch of rules
Week 2	10+ rules; AI output noticeably more like you
Month 1	30+ rules; style dimensions automatically calibrated
Ongoing	The rule library grows steadily, and new patterns are captured automatically