A Chinese-first
SKILL.mdfor AI coding agents. 让 Claude Code / Codex CLI / Cursor 用中文回复更短更密，不丢技术判断。
curl -fsSL https://raw.githubusercontent.com/Cpp1022/concise/main/install.sh | sh
# or: enable "every new chat starts concise" in addition to installing the skill
curl -fsSL https://raw.githubusercontent.com/Cpp1022/concise/main/install.sh | sh -s -- --default ultraIn chat:
/concise # enable ultra (default) for this chat
/concise lite # expression compression only
/concise ultra # expression + content compression
stop concise # back to normal
- Chinese-first: tuned for Chinese technical writing;
ultracuts 60-80% of non-code tokens in Chinese, 20-30% in English. - Preserves precision: code blocks, error output, command parameters, commit messages, and PR bodies stay verbatim.
- Auto-relax: backs off compression for safety warnings, multi-step procedures, parameter-heavy explanations, and ambiguous cases.
- Cross-agent: one
SKILL.md, three install targets — Codex CLI, Claude Code, Cursor.
Q: How is this different from a "be brief" system prompt? A: "Be brief" compresses wording but keeps the same output structure (intro / sections / recap). concise strips the structure too, and separates expression compression from content compression, so precision doesn't drop with filler.
Q: Will this make the agent think less? A: No. The skill only controls output style, not tool use or reasoning depth. TDD, debugging, and brainstorming workflows auto-relax when they need to expand.
Q: Does it conflict with my own Cursor rules? A: No. concise lives at the response-style layer; your project rules live at the code-style / architecture layer. Both can be on at the same time.
Q: Can I enable it only for specific chats?
A: Yes. Install only concise (not concise-default), then use /concise in the chats where you want it.
Suggested GitHub topics:
agent-skills
skill-md
codex
codex-cli
claude-code
cursor-rules
ai-coding-agent
concise-mode
developer-productivity
- Use
concisewhen you want the current chat to become concise. - Use
concise-defaultwhen you want every new chat to start concise.
concise controls response style. concise-default is a setup helper that installs that style into new-session instructions.
| Platform | Status | Notes |
|---|---|---|
| Codex CLI | Supported | Native SKILL.md install target. concise-default writes ~/.codex/instructions.md. |
| Claude Code | Supported as target | concise-default writes a managed block in ~/.claude/CLAUDE.md. The repo includes a Claude Code plugin manifest. |
| Cursor | Supported as rules target | concise-default writes ~/.cursor/rules/concise.mdc. Cursor uses Rules, not native SKILL.md. |
| OpenCode | SKILL.md-compatible, not yet a target | OpenCode can load SKILL.md from supported skill paths, but concise-default does not yet write OpenCode defaults. |
| GitHub Copilot / VS Code | SKILL.md-compatible, not yet a target | Copilot skills use .github/skills/ or ~/.copilot/skills/; concise-default does not yet write Copilot defaults. |
Windows:
mkdir $env:USERPROFILE\.codex\skills\concise -Force
copy concise\SKILL.md $env:USERPROFILE\.codex\skills\concise\SKILL.mdUnix-like:
mkdir -p ~/.codex/skills/concise
cp concise/SKILL.md ~/.codex/skills/concise/SKILL.mdInstall this only if you want a command that turns concise mode on for every new chat.
Windows:
mkdir $env:USERPROFILE\.codex\skills\concise-default -Force
copy concise-default\SKILL.md $env:USERPROFILE\.codex\skills\concise-default\SKILL.md
mkdir $env:USERPROFILE\.codex\bin -Force
copy concise-default\scripts\concise-default.py $env:USERPROFILE\.codex\bin\concise-default.py
copy concise-default\scripts\concise-default.cmd $env:USERPROFILE\.codex\bin\concise-default.cmdUnix-like:
mkdir -p ~/.codex/skills/concise-default ~/.codex/bin
cp concise-default/SKILL.md ~/.codex/skills/concise-default/SKILL.md
cp concise-default/scripts/concise-default.py ~/.codex/bin/concise-default.py
cp concise-default/scripts/concise-default.sh ~/.codex/bin/concise-default
chmod +x ~/.codex/bin/concise-defaultAdd ~/.codex/bin to PATH, or run the wrapper by full path.
These locations install the SKILL.md files for discovery. The concise-default command still needs the wrapper from concise-default/scripts/ on PATH if you want shell toggles.
Claude Code:
mkdir -p ~/.claude/skills
cp -r concise ~/.claude/skills/concise
cp -r concise-default ~/.claude/skills/concise-defaultOpenCode:
mkdir -p ~/.config/opencode/skills
cp -r concise ~/.config/opencode/skills/concise
cp -r concise-default ~/.config/opencode/skills/concise-defaultGitHub Copilot / VS Code:
mkdir -p ~/.copilot/skills
cp -r concise ~/.copilot/skills/concise
cp -r concise-default ~/.copilot/skills/concise-defaultProject-level Copilot skills can also live under .github/skills/.
Cursor does not load SKILL.md directly. Use concise-default on to generate ~/.cursor/rules/concise.mdc.
This repository includes .claude-plugin/plugin.json, so it can be used as a Claude Code plugin source or listed from a Claude Code plugin marketplace. For a marketplace listing, point the plugin source at this GitHub repository.
Use these in chat:
/concise
/concise lite
/concise ultra
stop concise
normal mode
Levels:
lite: expression compression onlyultra: expression compression + content filtering
Default: ultra.
Use these in shell:
concise-default on ultra
concise-default status
concise-default offValid levels: lite, ultra.
It makes agent responses shorter and denser while preserving technical accuracy and decision quality.
Use it when you want an agent to:
- use fewer tokens
- reduce filler and repetition
- keep conclusions first
- stay compact in Chinese technical writing
- avoid repeated conclusions and unasked-for expansion
ultra keeps only incremental information, leads with the conclusion, avoids restating the user's request, and expands only when the user asks for analysis, details, causes, or boundaries.
- Cursor:
~/.cursor/rules/concise.mdc - Claude Code: marked block in
~/.claude/CLAUDE.md - Codex CLI:
~/.codex/instructions.md - Saved level:
~/.config/concise/config
Codex detail: concise-default on writes the full concise skill body into ~/.codex/instructions.md. New Codex sessions splice that file into base_instructions. It does not change the current conversation retroactively.
concise-default status separates two Codex states:
Codex CLI: ON:~/.codex/instructions.mdexists.Codex App: ON: the most recent Codex session appears to include that instruction text.
This repo is structured for skill directories and marketplaces that accept SKILL.md-based agent skills.
Initial publishing targets:
- SkillsMD
- Skillzwave
- SKILLS.pub
- Bogen
After collecting usage feedback:
- OpenAI Skills Catalog
- LLMSkills
- Agent Skills Finder
- Cursor rules directories and awesome lists
- GitHub Copilot / VS Code awesome lists
When submitting, link to this GitHub repository as the source of truth. Marketplace pages should describe Codex CLI as the primary supported platform, with Claude Code and Cursor Rules as generated default targets.
The skill intentionally relaxes compression when clarity matters more:
- safety warnings
- irreversible operations
- multi-step instructions where order matters
- parameter-heavy or branch-heavy explanations
- repeated follow-up questions
- explicit requests for analysis or edge cases
Code blocks, commit messages, and PR bodies stay normal unless the user asks to compress them.
User:
Explain why this React component re-renders.
Normal:
This component is probably re-rendering because a new object reference is being created on each render, which causes React to treat the prop as changed.
Concise ultra:
inline obj prop -> new ref -> re-render. Wrap with useMemo.
.
|-- .claude-plugin/
| `-- plugin.json
|-- concise/
| `-- SKILL.md
|-- concise-default/
| |-- SKILL.md
| `-- scripts/
| |-- concise-default.py
| |-- concise-default.cmd
| `-- concise-default.sh
|-- README.md
|-- LICENSE
`-- .gitignore
MIT