sun-style-writing: a Chinese narrative skill that teaches models to stop explaining feelings

The problem sun-style-writing attacks: models that narrate feelings instead of events

Ask a general-purpose model to write a Chinese farewell scene and you get 百感交集, 青春, and 落幕 in the same paragraph. The README makes exactly this point with a two-version comparison: a generic model writes 走出公司大门的那一刻，我百感交集, while the sun-style version writes 前台把我的工牌剪断, 剪刀卡了两次, 她换了一把. The difference is not vocabulary. It is that the second version hands the reader two scissors and lets the reader supply the grief.

The project is aimed at a specific writer: someone with real material (a scene, a person, a date, a physical object) who wants prose that stays cold on the surface and still lands. The README lists the suitable inputs as true experiences with verifiable details, character stories, memoirs, reportage, and drafts that are already full of reflection and need half the sentences deleted. It is not for interface documentation, work reports, or anyone trying to manufacture literature out of line breaks. The README's own joke about this is blunt: 接口没有白月光.

Ten techniques and three gate rules: what the skill actually contains

The skill ships as a folder, not a runtime. The repository layout is three files: SKILL.md, VERSION, and references/style-anatomy.md. SKILL.md is what the model reads during normal use; style-anatomy.md is the breakdown manual the model returns to when it is unsure how a technique applies. VERSION exists so you know which release you installed.

The ten techniques are named in the README as 白描, 对照, 反复与变奏, 伏笔与照应, 物件叙事, 意象, 蒙太奇, 留白, 反高潮, and 回环, plus a half-technique called 称谓变化 (a shift in how a character is addressed, for example 妈妈 becoming 景甜 when the relationship is over). The README is explicit that these are not ten tasks to complete. 白描 appears every time; the rest depend on the material. The stated failure mode of over-application is that the characters have nowhere left to sit.

The three gate rules are the part that actually constrains output: if an action can speak, do not explain the psychology; if an object can speak, do not summarise the emotion; if an earlier passage can return once, do not write a new lyrical passage. A fourth rule applies to non-fiction only: do not invent numbers. The README's example is the eleven-second silence, which works because it really happened. If the writer cannot remember, the instruction is to write that they cannot remember.

Installing sun-style-writing into an agent's skills directory

The README gives no build step and no package manager. Installation is a message to an agent that supports Skills, pointing at the repository:

帮我安装这个 skill：https://github.com/KKKKhazix/sun-style-writing

The README states that a normally behaving agent will locate the sun-style-writing folder in the repository and install it. If the agent does not, the fallback is to download the .skill package from the Releases page, or to copy the sun-style-writing folder into the local skills directory at this path:

~/.agents/skills/sun-style-writing/

After installation, the README says nothing visible happens on screen. That is presented as expected rather than as a problem to debug.

To invoke it, name the skill directly, because the README warns that routing can fail:

使用 $sun-style-writing，把这段经历写成一篇克制但有后劲的叙事。

A plain-language invocation is also documented, asking for the 白描 and 留白 feel of 《我的女友景甜》, with the instruction not to be sentimental and to let actions, numbers, and objects speak.

Where it breaks: no numbers, no sea, and no substitute for having been there

The most consequential limitation is stated as a rule, not a bug: if the material contains no numbers, the skill will not add any, and if the material contains no sea, it will not dig one. For a writer who habitually relies on a model to smooth over gaps, this is the opposite of what they want. The model will stop at not knowing rather than guess, which the README frames as 不是漏写，是不替人物知道.

The second limitation is scope. The README excludes explanatory writing, API documentation, and weekly reports outright, on the grounds that there is no narrative material to work with. It also excludes requests to reproduce the original essay's plot, its fixed lines, or the voice of a real person, and it declines to answer whether the fifty million dollars should have been paid. The project describes itself as independent literary-technique research with no affiliation to 孙宇晨, 景甜, or the other people in the text, and it does not publish the full original essay. The reference file contains only short fragments for technique analysis.

The third limitation is maintenance of behaviour rather than code. The README has a feedback section for what it calls relapse: the model writing 那一刻，我终于明白了 at the exact moment it should stay silent. Issues are invited with the source material (redacted if needed), the model used, the offending passage, and the sentence where it started grieving on the writer's behalf. That is a prompt-behaviour problem, and no version number fixes it permanently.

Alternatives: house style guides and generic rewriting prompts

The nearest alternative is not another repository but the ordinary practice of pasting a style guide into a system prompt. A style guide lists preferences; sun-style-writing lists techniques plus prohibitions, and it separates the quick reference (SKILL.md) from the deeper anatomy (references/style-anatomy.md) so the model can load the second only when needed. Whether that two-file split changes output quality is not something the README claims, and there is no evaluation data in the repository to check it against.

A second alternative is a general-purpose rewriting prompt such as 让文字更克制. The difference in approach is that a generic instruction gives the model no vocabulary for what to do instead, so it tends to delete adjectives and stop there. sun-style-writing names ten specific moves, including 蒙太奇 (compress nineteen years into one cut but keep the eleven seconds intact) and 反高潮 (open with fifty million dollars, close on sixty-something strokes per minute). Named moves are easier for a model to apply selectively than an adjective like 克制.

A third alternative is fine-tuning on a corpus. That requires the corpus, which this project deliberately does not distribute, and it locks the behaviour into weights rather than a folder you can edit. The skill approach trades depth for editability: you can open SKILL.md and change a rule.

Licence, versioning, and what upgrading costs you

The project is MIT licensed, and the README adds its own note that the fifty million dollars is not distributed with the licence. MIT means you can copy, modify, and redistribute the skill files, including inside a commercial product, provided the licence text travels with them. That is a permissive arrangement, but it says nothing about the underlying essay: the README states that the full original text is not included and that only short fragments appear in the reference file for technique analysis. If you plan to redistribute the skill, that distinction is the one to check, and it is a question for your own legal reading rather than something this article can settle.

Upgrade cost is low by design. The folder carries a VERSION file, so you can tell which release is installed, and the current release is v1.0.0 from 2026-08-27. The last push to the repository was on 2026-08-27, which is recent. Because the payload is Markdown rather than code, an upgrade means replacing SKILL.md, style-anatomy.md, and VERSION; there is no dependency graph, no lockfile, and no migration step. The real cost of upgrading is re-reading the technique list, since the README notes that the ten techniques are a menu and not a checklist, and a new version could shift which one is always on.

Who should adopt sun-style-writing, and what to check first

Adopt it if you write Chinese narrative nonfiction from material you can verify, and your recurring problem is that drafts arrive pre-explained. The skill is small, MIT licensed, and installs by copying one folder, so the cost of trying it is a single agent session. Writers who already cut their own sentiment will get less from it, because the skill's main contribution is refusing to add what is not there.

Do not adopt it for technical documentation, status reports, or any task where the model is supposed to fill gaps in your knowledge. The non-fiction rule against inventing numbers makes it actively unsuitable for drafting anything that will be read as a factual record you did not personally verify.

Before relying on it, verify three things. Open sun-style-writing/SKILL.md and read the three gate rules in full, since they are what the model will actually enforce. Run one paragraph through it that contains a concrete action and a real date, and check whether the output preserves both. Then watch for the relapse sentence the README names, 那一刻，我终于明白了, because its appearance means the skill is loaded but not holding, and the repository's issue tracker is where that gets reported.

Editorial conclusion

Adopt sun-style-writing if you write Chinese narrative nonfiction from real, checkable material and you want the model to cut sentiment rather than add it. Do not adopt it for API docs, weekly reports, or any draft where you want the model to supply missing details, because the skill explicitly will not fill gaps and its non-fiction rule forbids inventing numbers. Before you commit, read SKILL.md and references/style-anatomy.md in the sun-style-writing folder, and confirm your agent actually loaded the skill by asking it to rewrite a paragraph that contains a concrete action and a date. If the output still contains a sentence like 那一刻，我终于明白了, the skill did not take.

Frequently asked questions

Official sources

1. Issues
2. KKKKhazix/sun-style-writing on GitHub
3. License: MIT
4. README
5. Releases