跪拜 Guibai
← All articles
Frontend · Trae

A Four-Skill TRAE Work Pipeline That Writes Release Notes in Two Minutes

By 不如摸鱼去 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The workflow treats AI output as a first draft to be processed, not a final product to be tweaked, which sidesteps the burnout of editing generic LLM prose. Developers who publish on platforms with idiosyncratic formatting rules, like WeChat, get a concrete pattern for chaining small, single-purpose skills instead of wrestling with one giant prompt.

Summary

Asking an LLM to produce a publishable article in one shot usually yields grammatically correct filler that sounds nothing like the author. The `wot-ui` maintainer behind this workflow stopped treating AI as a ghostwriter and started treating it as a drafting and editing pipeline inside TRAE Work. Four custom skills handle distinct stages: `humanizer-zh` purges stock phrases and hollow conclusions from a rough draft, `writing-style` reworks the text to match the author's own cadence and vocabulary, `style-extractor` reverse-engineers structural techniques from articles worth learning from, and `wechat-link-formatter` converts Markdown links into the plaintext format WeChat public accounts require. The author supplies the raw material — cases, opinions, real experience — and the pipeline assembles, de-fluffs, and polishes it. Release announcements that once took half a day now finish in about two minutes, with the remaining time going back to actual thinking and building.

Takeaways
Treating AI output as a rough first draft, rather than a near-final article, removes the pressure to get a perfect result from a single prompt.
`humanizer-zh` scans for and removes common Chinese AI-writing tics: empty background paragraphs, forced three-point lists, and formulaic conclusions.
`writing-style` is built by feeding several of the author's own past articles into TRAE Work so it can learn sentence rhythm, paragraph structure, and habitual phrasing.
`style-extractor` analyzes admired articles to identify reusable techniques — such as how an opening cuts padding or how code and explanation are interleaved — without copying the source's voice.
`wechat-link-formatter` automates the tedious conversion of Markdown hyperlinks into the "text + URL" format required by the WeChat public account editor.
Release announcements that previously required half a day of drafting and editing now take roughly two minutes with the four-skill pipeline.
Conclusions

The core shift is lowering expectations: the author stopped asking the model to write a finished piece and instead asked it to organize supplied material into a draft, which made the output immediately more useful.

Breaking writing into discrete, toolchain-style steps mirrors how developers already work with code — linting, formatting, refactoring — and applies that same sensibility to prose.

The `style-extractor` skill is used for study, not imitation; the author cherry-picks structural techniques and adds them to `writing-style`, which keeps the output from drifting toward someone else's voice.

Concepts & terms
TRAE Work
An AI-powered workspace that supports chaining custom skills to automate multi-step tasks such as drafting, editing, and formatting documents.
humanizer-zh
A custom skill for Chinese text that detects and removes AI-generated clichés, hollow rhetoric, and overly rigid sentence structures while preserving the original information and opinions.
writing-style
A custom skill trained on an individual's previously published articles to replicate their specific tone, paragraph rhythm, and habitual expressions when rewriting drafts.
style-extractor
A custom skill that analyzes a given article's opening strategy, content structure, paragraph flow, and signature phrases so a writer can adopt specific techniques without copying the source's voice.
wechat-link-formatter
A custom skill that converts Markdown hyperlinks into plain "text + URL" pairs to meet the formatting constraints of the WeChat public account platform.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗