Cli Design

CLI Design

Use this skill to make command-line programs feel humane, predictable, and composable. Ground decisions in the Command Line Interface Guidelines at https://clig.dev/: CLIs are text UIs for humans first, but they still need to behave as good parts in scripts, pipelines, CI, and larger automation systems.

The goal is not to copy old Unix behavior blindly. Prefer conventions because they make interfaces guessable; break them only when doing so clearly improves usability.

Operating model

When designing or reviewing a CLI, separate the work into four surfaces:

1. Command shape — names, subcommands, positional args, flags, defaults.
2. Conversation — help, errors, suggestions, prompts, confirmations, next

steps.

1. I/O contract — stdout, stderr, stdin, exit codes, TTY behavior,

machine-readable modes.

1. Safety and evolution — destructive actions, dry-runs, idempotence,

config, env vars, deprecation.

A good answer should explicitly cover all four when the task is broad. For a narrow task, focus on the relevant surface but still check for knock-on effects.

First-pass questions

Ask only the questions needed to avoid designing the wrong interface. If the user's intent is already clear from code or context, inspect first and proceed.

• Who is the primary user: humans in a terminal, scripts/CI, or both?
• Is this a single-purpose command or a git-style multitool with subcommands?
• What is the primary output: human status, transformed data, files, remote

state, or side effects?

• Which operations are destructive, expensive, remote, or surprising?
• What should be stable for automation: command names, flags, JSON schema, plain

text, config, env vars?

Design principles

Human-first, script-friendly

Optimize default output for a human reading a terminal. Preserve scriptability with explicit stable modes such as --json, --plain, or --quiet rather than making the default terse and cryptic.

Say just enough

A CLI says too little when users wonder whether it is hung. It says too much when important information is buried in noisy logs. Show progress and next steps for long or multi-step operations, but keep final output scannable.

Be conversational

Treat each invocation as part of a conversation. Good CLIs help users recover:

• explain what went wrong in ordinary language
• show the invalid value or missing requirement
• suggest the likely correction or next command
• avoid raw stack traces unless a debug mode was requested

Be robust and feel robust

Make unexpected input graceful. Prefer idempotent operations where practical. Keep users informed during slow work. A CLI feels fragile when it hangs silently, prints alarming internals, or makes state-changing guesses without consent.

Command and flag design

Prefer an argument parsing library for the implementation language. It should handle parsing, help, validation, shell completion where possible, and spelling suggestions.

Design rules:

• Use clear verbs for subcommands: init, list, show, create, delete,

sync, login.

• Prefer standard flag names when a convention exists: --help, --version,

--verbose, --quiet, --output, --format, --json, --force, --dry-run, --config.

• Provide long versions for all flags. Use short flags only for common actions;

do not exhaust the namespace with obscure one-letter flags.

• Prefer flags over positional args for anything optional, ambiguous, or likely

to grow.

• Keep positional args for required primary nouns where order is obvious:

cp .

• Avoid catch-all subcommands or ambiguous magic if the same behavior can be

expressed as clear flags/subcommands.

• Keep changes additive where possible. A flag, env var, config key, command

name, and output schema are all public interfaces once released.

When proposing a CLI shape, include a small command synopsis and 2-4 representative examples.

Help and discoverability

Every command and subcommand should support -h and --help. For multitool CLIs, also support help and help when feasible.

Top-level help should include:

• one-sentence purpose
• usage synopsis
• the most common commands or flags first
• 1-3 examples that mirror real tasks
• where to find full documentation or support

If a command needs arguments and is run without them, show concise help rather than a bare error. Concise help should include what the command does, one or two examples, common flags, and a pointer to --help.

Lead with examples because users copy examples before they read reference text. For complex workflows, tell a short story through examples that progress from simple to advanced.

Output contract

Follow the stream contract:

• Primary command output goes to stdout.
• Machine-readable output goes to stdout.
• Errors, warnings, progress, logs, and prompts go to stderr.
• Success exits 0; failure exits non-zero. Map distinct non-zero codes only

when callers can act on them.

TTY behavior:

• If stdout is a TTY, optimize for humans: headings, spacing, aligned columns,

color with restraint.

• If stdout is not a TTY, disable decoration that would break piping unless

explicitly requested.

• Disable color when not writing to a terminal, and respect common controls such

as NO_COLOR where appropriate.

• If a command expects piped input but stdin is interactive, print help or an

explanatory message instead of hanging silently.

Machine-readable modes:

• Offer --json for structured automation when the data has nested shape or

stable fields matter.

• Offer --plain or a simple tabular line format when users need grep, awk,

or shell-friendly output.

• Keep JSON schemas stable. If human output changes, tell users to script

against --json/--plain rather than default text.

Errors and recovery

Good error messages have four parts when possible:

1. What failed.
2. Why it failed.
3. What the user can do next.
4. A concrete command or flag to try.

Example pattern:

Error: could not find app "billing-api" in project "prod".

Run `tool apps list --project prod` to see available apps.

Do not silently autocorrect state-changing input. If the likely fix is safe, suggest it. If it mutates state, ask interactively or require an explicit flag.

Safety for destructive operations

Destructive or expensive actions need friction proportional to risk.

• Mild local change: prompt interactively or require --force in

non-interactive mode.

• Broad or remote change: support --dry-run and show what will change before

applying it.

• Severe deletion or irreversible action: require typing the resource name, or

support a scriptable --confirm flag.

Never make prompts the only path. Every interactive flow needs a non-interactive way to pass the same information via args, flags, config, env vars, or stdin.

Interactivity and automation

Prompt for missing input only when stdin is interactive. In scripts, CI, or piped use, fail with a clear message explaining the missing flag or value.

For multi-step workflows, print the next useful command after success. This teaches the interface and makes state visible:

Created project demo.

Next: tool deploy demo --env staging

Support --verbose for more detail and --quiet for suppressing non-essential messaging. Keep errors visible even in quiet mode unless the user explicitly redirects them.

Configuration and environment

Use config files and environment variables for values that are stable across invocations, such as auth tokens, default org/project, region, or output preference.

Make precedence explicit in docs and help when relevant. A common order is:

1. command-line flags
2. environment variables
3. project-local config
4. user config
5. defaults

Do not hide surprising behavior in config. If a command will affect remote state or many files because of config, show the resolved target before acting or in --dry-run.

Review checklist

When reviewing an existing CLI, report findings under these headings:

1. Top issues — the highest-impact usability or automation problems.
2. Command shape — names, args, flags, defaults, consistency.
3. Help and discovery — help text, examples, suggestions, docs links.
4. I/O and scriptability — stdout/stderr, exit codes, TTY behavior,

JSON/plain modes.

1. Errors and safety — recovery messages, prompts, destructive actions,

dry-runs.

1. Recommended changes — concrete command examples or implementation notes.

Prefer specific rewrites over vague advice. For example, write the improved error text, proposed help section, or new command synopsis.

Implementation checklist

Before calling a CLI design complete, verify:

• cmd --help and cmd -h work.
• Each subcommand has help and examples.
• No-arg behavior is helpful: concise help or intentional interactive mode.
• stdout contains only primary output; stderr contains logs/progress/errors.
• exit code is 0 on success and non-zero on failure.
• color/progress adapts to TTY and can be disabled.
• machine-readable output exists where users will script against results.
• destructive actions have confirmation, --force, --confirm, or --dry-run

as appropriate.

• prompts are never required for automation.
• errors explain what happened and how to recover.
• docs state stability guarantees and deprecation path for interfaces.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: markphelps
• Source: markphelps/agent-plugins
• License: MIT

Install and usage instructions live in the source repository linked above.