No AI Style
Natural writing. Faithful meaning.
An agent skill and DeepSeek Harness plugin for academic and professional writing
English ·
中文 ·
日本語
---
## What it does
It removes the phrasing that makes a text read like a template — the throat-clearing, the stacked hedges, the self-justification — and leaves the claims alone. Facts, numbers, citations, methods and limitations survive as written, and the voice stays yours.
> **Before** — For clarity, it should be mentioned that the device supports two USB ports.
>
> **After** — The device supports two USB ports.
## The three modes
| Mode | Name | What you get |
| :-- | :-- | :-- |
| `audit` | **Check** | Findings and suggestions only. Ask whether it is ready to submit and you also get a verdict and the unresolved issues. The source file is never touched. |
| `revise` | **Polish** | An edited version within the scope you asked for. |
| `compose` | **Draft** | A draft built from the notes and evidence you supplied. |
Light editing is the default; ask for a stronger rewrite when you want one. The skill follows the language of the text and leaves passages that already work untouched.
```text
Use no-ai-style to check this.
Use no-ai-style to polish it.
Use no-ai-style to draft from these notes.
```
Two markers keep it honest: `[EVIDENCE GAP]` flags a claim that needs a source and blocks a `ready` verdict, and `[INTERPRETATION FLAG]` flags a reading the author should confirm.
## In the composer
The DSH plugin adds a ⭐ **no ai style** button to the input row. Pick a mode and the matching instruction lands at the front of your draft — in the interface language, Chinese, English or Japanese — with whatever you already typed left intact.
Insertion is undoable. A draft that carries file or command references is refused rather than flattened, so those references survive.
## Install
### DeepSeek Harness
From the app: **Settings → Plugins** → add `github:01NGVI-QINGYI/no-ai-style`, then restart.
From the CLI — `--profile` is the profile you actually run, `web` for the `dsh web` build and `desktop` for the desktop app:
```bash
dsh plugin --profile web add github:01NGVI-QINGYI/no-ai-style
```
The plugin registers its own skill provider, so the skill reaches every session in that profile without touching your project, custom or user skill roots.
Prefer the skill on its own, without the plugin?
Run this inside your DSH project; it installs to `.agents/skills`:
```bash
npx skills add 01NGVI-QINGYI/no-ai-style --agent universal --copy
```
### Any other agent
```bash
npx skills add 01NGVI-QINGYI/no-ai-style
```
Choose your agent when prompted. Omit `--global` for a project install.
Per-agent commands
| Agent | Command |
| :-- | :-- |
| Codex | `npx skills add 01NGVI-QINGYI/no-ai-style --global --agent codex` |
| Claude Code | `npx skills add 01NGVI-QINGYI/no-ai-style --global --agent claude-code` |
| Cursor | `npx skills add 01NGVI-QINGYI/no-ai-style --global --agent cursor` |
| Pi | `npx skills add 01NGVI-QINGYI/no-ai-style --global --agent pi` |
To install by hand, copy `skills/no-ai-style/` into your agent's skills directory.
## Optional checks
Two local scripts, Python 3.9+, no dependencies. Run them from the repository root:
```bash
python3 skills/no-ai-style/scripts/audit_prose.py draft.md --lang auto
python3 skills/no-ai-style/scripts/compare_anchors.py before.md after.md --format json
```
Neither uploads text. Findings still need a human read: an unclosed comment or code fence produces a warning, and `--strict` turns missing-evidence markers or unclosed markup into exit code 1.
## License
[MIT](LICENSE) · skill v1.0 · DSH plugin v1.2.2 · by [01NGVI-QINGYI](https://github.com/01NGVI-QINGYI)
The goal is readable, faithful prose, not AI-detector evasion. Full rules are in [SKILL.md](skills/no-ai-style/SKILL.md).