agentic-cli-design（agentic CLI 设计）

Agentic CLI Design

Overview

Agentic CLI Design is a design philosophy that treats CLIs not as "tools for humans to operate in terminals," but as protocols/APIs that LLMs/agents can safely, reliably, and repeatably invoke.

The success criteria is that agents can operate the CLI:

• Without confusion (options are explicit)
• Without breaking things (safe-by-default)
• Without getting stuck (non-interactive, timeout/retry strategies)
• Repeatedly (idempotent, re-execution tolerant)
• With self-healing (observable, self-describing)

This skill provides a comprehensive framework for designing and evaluating CLI tools for agent use.

Core Definition

Agentic CLI Design = *Design principles for CLIs that enable LLM/agents to execute commands safely and reliably in non-interactive, repetitive, failure-prone environments.*

Rather than optimizing for human "feel," we optimize for machines to read, decide, re-execute, and recover.

The 7 Principles

P1. Machine-readable (Primary Output Format)

Machine-readable output is the primary interface, not human-readable text.

• Provide --json / --output json|yaml|text as a first-class option
• Strict separation: stdout=results / stderr=logs/progress (never mix)
• Errors are also structured (preferably JSON)
• Schema stability (manage breaking changes with schemaVersion)

When to use: Every CLI command should support structured output.

P2. Non-interactive by Default

Interactive prompts should not be the default behavior.

• Don't assume interactive prompts
• Provide --yes / --force / --no-confirm / --non-interactive flags
• Must complete successfully in TTY-less environments (CI, job runners)

When to use: Any command that might prompt for user input.

P3. Idempotent & Replayable

Commands should be safe to run multiple times with the same result.

• Send/create operations accept dedupe-key / client-request-id
• "Already exists" behavior is configurable: --if-exists skip|update|error
• Retrieval uses explicit pagination: --limit --cursor --all

When to use: Any write operation or data retrieval command.

P4. Safe-by-default

Destructive operations require explicit confirmation.

• Destructive operations support --dry-run / --confirm <id>
• Deletion requires --force or similar (default prevents accidents)
• Permissions/scopes are minimized; insufficient permissions return "next steps"

When to use: Any command that modifies or deletes data.

P5. Observable & Debuggable

Operations must be traceable and debuggable.

• Provide --verbose / --debug / --log-format json
• Accept --trace-id for correlation IDs
• Exit codes are categorized for automatic recovery:

- 0: success - 2: invalid arguments - 3: authentication/permission - 4: retryable (rate limit, transient errors)

When to use: All commands should support observability flags.

P6. Context-efficient

Minimize token/context consumption for LLM agents.

• --fields/--select (projection)
• --output ndjson (streaming)
• Default is summary; details via get/--include-* explicitly
• Rich server-side filtering (since/until/query/type...)

When to use: Commands that return large datasets or detailed information.

P7. Introspectable (Self-describing CLI)

The CLI itself can emit its specification in machine-readable format.

MCP tools have schema definitions, but CLIs are often black boxes. Solution: CLI emits its own specification.

• commands --json (list commands and arguments)
• schema --command... --output json-schema (per-command JSON Schema)
• --help --json (examples, exit codes, error vocabulary)
• install-skills (emit/install the CLI's own agent skill from <project dir>/skills)
• --output json top-level fixed fields:

- schemaVersion, type, ok

Recommended install-skills behavior:

• Installs the CLI's own usage skill, not arbitrary third-party skills
• Reads skill source from <project dir>/skills
• Default destination: ./.agents/skills
• install-skills --global: ~/.agents/skills
• install-skills --claude: ./.claude/skills
• install-skills --claude --global: ~/.claude/skills

When to use: For CLIs that need to be discoverable and self-documenting to agents.

Reference Documentation

This skill includes detailed reference materials organized by topic:

references/principles.md

Detailed explanation of each of the 7 principles with concrete examples, implementation guidance, and best practices.

Use when: You need in-depth understanding of a specific principle or implementation examples.

references/scorecard.md

Complete checklist for reviewing CLI design against the 7 principles. Each principle has specific checkboxes for evaluation.

Use when: Reviewing existing CLIs for agent-friendliness or validating new CLI designs.

references/templates.md

Concrete templates for:

• JSON response formats (success/failure)
• Exit code conventions
• Introspection command patterns
• Authentication status responses

Use when: Implementing specific features like JSON output, error handling, or self-description.

references/anti-patterns.md

Common failure patterns that break agent compatibility, including:

• Mixed stdout/stderr
• Inconsistent JSON schemas
• Interactive-by-default behavior
• Unsafe destructive operations
• Context explosion
• Browser-only authentication

Use when: Troubleshooting agent failures or avoiding common pitfalls.

Quick Start

For New CLI Design

1. Review the 7 principles above
2. Use references/scorecard.md as a design checklist
3. Implement using patterns from references/templates.md
4. Avoid patterns in references/anti-patterns.md

For Existing CLI Review

1. Run through references/scorecard.md
2. Identify gaps in principle coverage
3. Check references/anti-patterns.md for current issues
4. Prioritize fixes based on agent use cases

For Agent Skill Creation

When creating an Agent Skill for a CLI tool, include:

• Task-specific recipes (minimal command sequences)
• Guardrails (dry-run → confirm → execute patterns)
• Input/output types (typical success/failure JSON)
• Error recovery procedures (rate limit, auth_required, etc.)
• Recommended defaults (--json/--non-interactive/--limit)
• An install-skills flow so the CLI can install its own skill from <project dir>/skills

Authentication Considerations

For OAuth/headless authentication:

• Prefer Device Authorization Grant (RFC 8628) as first choice
• Provide auth status --json for agents to check prerequisites
• Support auth export / auth import for headless environment migration
• In --non-interactive mode: return error + next steps (don't prompt)

See references/templates.md for authentication response formats.

Scoring Your CLI

Each principle can be scored 0/1/2 points for easy comparison:

• 0: Not implemented
• 1: Partially implemented
• 2: Fully implemented

Maximum score: 14 points (2 points × 7 principles)

Use references/scorecard.md for detailed scoring criteria.

Related Concepts

• MCP (Model Context Protocol): Provides tool definitions with schemas; Agentic CLI Design makes standalone CLIs equally discoverable
• Agent Skills: Documentation layer that teaches agents safe CLI usage patterns
• Infrastructure as Code: Shares idempotency and declarative principles
• 12-Factor Apps: Shares configuration and observability principles

Examples in the Wild

Well-designed CLIs that demonstrate these principles:

• GitHub CLI (gh): Excellent --json support, non-interactive flags
• kubectl: Strong idempotency, declarative apply, dry-run support
• AWS CLI: Comprehensive --output json, pagination, filtering

When to Use This Skill

Use this skill when:

1. Designing a new CLI intended for agent/automation use
2. Improving an existing CLI to be more agent-friendly
3. Creating an Agent Skill for a CLI tool
4. Debugging agent failures with CLI tools
5. Reviewing CLI design for production readiness
6. Choosing between CLI tools for agent workflows

When NOT to Use This Skill

This skill may be overkill for:

• One-off scripts for personal use
• CLIs exclusively for human interactive use
• Prototypes not intended for production
• Tools with inherently interactive workflows (text editors, REPLs)

Further Reading

• references/principles.md - Deep dive into each principle
• references/scorecard.md - Evaluation checklist
• references/templates.md - Implementation templates
• references/anti-patterns.md - What to avoid