Best practices for inclusive CLIs

Problematic patterns

The “Getting Started Experience” section of Lucas’ article has a GIF video of a CLI utility printing --help output, featuring a decorative border.

						$ tool --help

╔ Getting started ════════════════════════════════════════╗
║                                                         ║
║  To scaffold a new project, run:                        ║
║                                                         ║
║    $ mytool init <directory>                            ║
║                                                         ║
║  If you already have a project set up and would like to ║
║  add, remove, or update its structure, run:             ║
║                                                         ║
║    $ mytool manage                                      ║
║                                                         ║
╚═════════════════════════════════════════════════════════╝

Usage: tool <command> [options]
Commands:
  tool init [directory]	creates a new project
  tool manage	allows you to manage an existing project

Options:
  --help	Show help	[boolean]
  --version	Show version number	[boolean]

1. Borders in TUIs should always be drawn with characters specifically intended for textual interfaces (e.g., boxdraw characters). While I do think the GIF followed this advice, I think it’s worth explicitly saying it. Accessible terminal emulators can figure out what these mean and factor them into what they report through an accessibility API. But breaking these borders up with descriptive text makes detection of readable text error-prone.

2. Borders should be used sparingly, as they end up causing issues when the window is re-sized.^note 1 Re-sizing terminal windows is quite common: think about the combined user-base of tiling window managers, tiling terminal session managers (Tmux, Screen, etc.), multiplexing terminal emulators, and plain old split-windows.

3. Decorative content in CLI output should be limited, since the output of CLI utilities can be piped through other programs. At the very least, these tools should be able to detect whether their standard output is being re-directed or piped and sanitize output accordingly.

The “Colours, Emojis, and Layouting” (sic) section has similar issues:

1. Nearly all animated spinners are extremely problematic for screenreaders. A simple progress meter and/or numeric percentage combined with flags to enable/disable them is preferable.

2. Excessive animation and color can harm users with attention and/or vestibular disorders, and some on the autism spectrum. Many tools offer a --color[=WHEN] flag where WHEN is always, never, or auto. Expecting users to learn all the color configurations for all their tools is unrealistic; tools should respect the NO_COLOR environment variable.

Recommen­dations

This is a non-exhaustive list of simple, baseline recommendations for designing CLI utilities.

Accessibility

1. Send your tool’s output through a program like espeak-ng and listen to it. Can you make sense of the output?

2. Refer to the latest WCAG publication (currently WCAG 2.2) and take a look at the applicable criteria. Many have accompanying techniques for plain-text interfaces.. Avoiding reliance on color in favor of whitespace and/or indentation for pseudo-headings are two sample recommendations from the WCAG.

3. Make sure your web-based documentation and forge frontends are accessible, or are mirrored somewhere with good accessibility. I love what the Gitea folks are doing, but sadly their web frontend has a number of critical issues.^note 2 For now, if your forge has accessibility issues, mirroring to GitHub and/or Sourcehut seems like a good option.

4. Avoid ASCII-art, and use presentational text sparingly. Include a way to configure output to be friendly to screen-readers and log files alike (if it isn’t already). For example, a simplified output mode can occasionally log a percentage-complete instead of a progress bar.

Familiarity

1. How “unique” is your tool’s output? Output should look as similar to other common utilities as possible, to reduce the learning curve. Keep it boring.

2. Follow convention: use POSIX-like options. Consider supplementing them with GNU-style long options if your tool has a significant number of them.^note 3

3. Avoid breaking changes to you program’s CLI. Remember that its argument parsing is an API, unless documentation explicitly states otherwise.^note 4 Semantic versioning is your friend.

4. Be predictable. Users expect git log to print a commit log. Users do not expect git log to make network connections, write something to their filesystem, etc. Try to only perform the minimum functionality suggested by the command. Naturally, this disqualifies opt-out telemetry.

Documentation

1. Write man pages! Man pages have a standardized,^note 5 predictable, searchable format. Many screen-reader users actually have special scripts to make it easy to read man pages. A man page is also trivial to convert to HTML for people who prefer web-based documentation.^note 6 If your utility has a config file with special syntax or vocabulary, write a dedicated man page for it in section 5 and mention it in a “SEE ALSO” section.^note 7

2. Try adding shell completions for your program, so users can tab-complete options. This is particularly helpful in shells like Zsh that support help-text in tab completions, especially when combined with plugins like fzf-tab that enable fuzzy-searching help-text (see code snippet 2).

3. Related to no. 2: use a well-understood format for -h and --help output. This makes auto-generating shell completions much easier. A great example is the Busybox coreutils’ help output: it is much more concise than man pages, but descriptive enough to serve as a quick reference. Alternatively, delegate the generation of both to a library that follows this advice.

4. Ensure that the whatis and apropos commands work as intended after installing your man pages. These commands parse the beginnings of man pages to give one-line summaries of programs, and often power advanced tab-completion setups.

						$ moac -
> -p
  9/11 (0)
  -P  -- power available to the computer (W)
> -p  -- password to analyze
  -s  -- password entropy
  -h  -- display this help message
  -r  -- interactively enter a password in the terminal; overrides -p
  -T  -- temperature of the system (K)
  -m  -- mass at attacker's disposal (kg)
  -q  -- account for quantum computers using Grover's algorithm

Miscellaneous

1. Either delegate output wrapping to the terminal, or detect the number of columns and format output to fit. Prefer the former when given a choice, especially when the output is not a TTY.

2. Be safe. If a tool makes irreversible changes to the outside environment, add a --dry-run or equivalent option.

3. If your tool has color output: disable color when the output is not a TTY, unless the user explicitly force-enables color via a command-line flag. Many tools support a --color flag that accepts the values “always”, “never”, and “auto”.

More opinionated considerations

These considerations are far more subjective, debatable, and deserving of skepticism than the previous recommendations. There’s a reason I call this section “considerations”, not “recommendations”. Exceptions abound; I’m here to present information, not to think on your behalf.

1. Remember that users aren’t always at their best when they read --help output; they could be trying to solve a frustrating problem, feeling a great deal of anxiety. Keep the output clean, predictable, boring, and fast. A 2-second delay and spinning fans will probably be extremely unpleasant for already-stressed users, especially if they need to use it often.^note 8

2. Include example usage in your man pages and accompanying documentation. Consider submitting the example usage to the tldr pages project if your tool gets popular.

3. Include an extended list of example command invocations and expected output. Make that document double as a test suite. My moac testdata and moac-pwgen testdata scripts are good examples. This can serve as a check for API stability, and even as a source of documentation.

4. Make your man pages as similar to other man pages on the target OS as possible. Many programs parse man pages, and expect them to follow a predictable structure. Try testing your man pages in multiple programs, just as people test Web pages in multiple browser engines. Some examples:

   • w3mman (included in w3m) is a good example to make sure auto-hyperlinking works.

   • Editors like Vim support looking up man pages for the currently-selected word. Try pressing Shift+k while your caret is on a command name.

   • Pandoc is another tool worth testing; it can convert man pages to a variety of different formats.

5. Conform to tools that share a similar niche. If you’re using Rust to make a fast alternative to popular coreutils: model its behavior, help-text, and man pages after ripgrep and fd. If you’re making a linter for Go: copy go vet.

6. If you want to keep your tool simple, make the output readable to both humans and machines; it should work well when streamed to another program’s standard input and when parsed by a person. This is especially useful when people redirect output streams to log files, and to screen readers.

7. Consider splitting related functionality between many executables (the UNIX way) and/or sub-commands (like Git). I split MOAC’s functionality across both moac and moac-pwgen, and gave moac three subcommands. The “Consistent commands trees” section of Lucas’ article has good advice.

8. Don’t conflate CLIs and TUIs. A CLI should be non-interactive; a TUI should be interactive. Exceptions exist for really simple interfaces (e.g. Magic-Wormhole and others like it) that accept user input; however, as the interface grows more complex, consider splitting the program into two sibling programs, one of which can have a “pure” non-interactive CLI.

9. Go above and beyond by writing separate integrations for environments like Emacspeak.^note 9

Name conflicts

This section might be the most important part of this post. If a CLI executable has a binary name conflict, packagers may have to re-name it. Otherwise, users will have to juggle $PATH overrides.^note 10

Before publishing your software, test for binary name conflicts. Many package managers have built-in functionality to search for package files. I recommend doing so with large repositories to test for conflicts.

						$ dnf provides /usr/bin/p
Last metadata expiration check: 0:48:19 ago [...]
perl-App-p-0.0400-19.fc36.noarch : Steroids for your perl one-liners
Repo        : fedora
Matched from:
Filename    : /usr/bin/p

$ curl https://cht.sh/fly
# fly
# Command-line tool for concourse-ci.
# More information: <https://concourse-ci.org/fly.html>.
[...]

Thanks to Emily for reminding me of name conflicts.