Agent Speak Readability — Development skill for Claude Code

agent-speak-readability

An agent skill that makes an agent's natural-language output readable to a human reader. The skill holds cross-language rules at the top level, Chinese rules under `lang/zh/`, and English rules under `lang/en/`. The core of the skill is a pre-send check written as edit steps, not as a quality to aim at.

The problem this skill fixes

An agent writes for itself. The agent already knows the actor, the object, the file name and the result, so the agent compresses all of them out of the sentence. The human reader does not know any of them. The text that comes out looks confident and carries almost no information.

This repository is not a style preference. This repository is a list of edits that an agent performs on a draft before the agent sends the draft.

Directory tree

agent-speak-readability/
├── README.md                      this file
├── CHANGELOG.md                   version history
├── LICENSE                        Apache License 2.0
├── SKILL.md                       skill entry point: 9 cross-language rules + the pre-send check
├── AGENTS.md                      repository-wide constraints for agents
├── CLAUDE.md                      symlink to AGENTS.md
├── docs/
│   └── suggested-claude-md-section.md   a replacement section for a user's global CLAUDE.md
└── lang/
    ├── zh/
    │   ├── PRINCIPLES.md          Chinese rules Z1 to Z10
    │   ├── ACCEPTANCE-CASES.md    four hand-written corrections, kept word for word
    │   └── NEGATIVE-SAMPLES.md    real Chinese failure samples from one session
    └── en/
        └── PRINCIPLES.md          English rules E1 to E9

How an agent uses the skill

1. Read SKILL.md and take the nine cross-language rules.
2. Read the file for the language you are about to write in. Chinese output uses lang/zh/PRINCIPLES.md. English output uses lang/en/PRINCIPLES.md.
3. Write the draft.
4. Run the ten check steps in section 4 of SKILL.md on the draft.
5. Run the extra chec