Design CLIs that AI agents can actually use

One CLI · Three Audiences · Three Channels

A skill that evaluates whether a CLI is reliably usable by AI agents and helps you design CLIs that serve humans, agents, and orchestration systems at the same time. Built around seven principles, a 14-criterion rubric, and a structured refactor playbook.

What it does

Evaluates whether an existing CLI is reliably usable by AI agents
Designs CLI interfaces that serve humans, agents, and orchestration systems simultaneously
Converts REST APIs and SDKs into agent-native CLI command trees
Reviews stdout contracts, exit code semantics, and error envelope design
Designs schema-driven self-description, dry-run previews, and schema introspection
Defines safety tiers (open / warned / hidden) for graduated command visibility
Designs delegated authentication so agents never own the auth lifecycle
Produces prioritized refactor plans with concrete interface examples

Documentation

Doc	What's inside
docs/install.md	Per-platform install (Claude Code / OpenClaw / Hermes / pi-mono / Codex / SkillsMP) and path summary
docs/changelog.md	Version history from v1.1.0 through v1.3.3
SKILL.md	Workflow guide loaded by the agent
references/	On-demand reference material — design patterns, rubric, checklists, examples, testing recipes, citations

Multi-Platform Support

The core SKILL.md is portable, and this repository includes metadata for the platforms listed below:

Platform	Status	Details
Claude Code	Full support	Native SKILL.md format
OpenClaw / ClawHub	Full support	`metadata.openclaw` namespace
Hermes Agent	Full support	`metadata.hermes` namespace, category: engineering
pi-mono	Full support	`metadata.pimo` namespace
OpenAI Codex	Full support	`agents/openai.yaml` sidecar
SkillsMP	Indexed	GitHub topics configured

Comparison: with vs. without this skill

Capability	Native agent	This skill
Evaluate whether a CLI is agent-native	No	Yes — structured diagnosis across 7 principles
Design stdout JSON contract	Inconsistent	Always — stable envelope with `ok`, `data`, `error`
Define exit code semantics	Ad hoc	Yes — documented, deterministic per failure class
Design layered `--help` and schema introspection	No	Yes — full self-description pattern
Design dry-run previews	Rarely	Always — request shape preview without execution
Define safety tiers for commands	No	Yes — open / warned / hidden tiers
Design delegated authentication	No	Yes — human manages auth lifecycle; agent uses token
Separate trust levels for env vs. CLI args	No	Yes — directional trust model
Produce prioritized refactor plan	Rarely	Always — P0 / P1 / P2 with examples
Score CLI across 14-criterion rubric	No	Yes — 0–2 per criterion with verdict

When to use

Evaluating whether an existing CLI is usable by an AI agent
Designing a new CLI interface for an API or SDK
Refactoring a human-first CLI to be machine-readable
Reviewing stdout, stderr, and exit code contract design
Defining dry-run, schema introspection, and self-description layers
Designing auth delegation and trust boundaries for agent safety
Producing a SKILL.md or skill docs from a CLI schema

Installation

See docs/install.md for per-platform install commands (Claude Code, OpenClaw / ClawHub, Hermes, pi-mono, OpenAI Codex, SkillsMP) and the installation paths summary.

Support

If this skill helps your work, consider supporting the author:

WeChat Pay

Alipay

Buy Me a Coffee

Give a Reward

Author

Agents365-ai

Bilibili: https://space.bilibili.com/441831884
GitHub: https://github.com/Agents365-ai

License

MIT

agent-native-design

Design CLIs that AI agents can actually use

What it does

Documentation

Multi-Platform Support

Comparison: with vs. without this skill

When to use

Installation

Support

Author

License

Similar Packages

Repository

Tags

Original Author

Publisher