TUI Design
SkillMediaDesign terminal user interfaces and interactive CLIs that stay usable, accessible, and scriptable. Use when building or reviewing a TUI, a full-screen terminal app, or an interactive command-line tool.
Available today. Use it from your connected AI after setup.
No other account needed.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
Then ask your AI: use the TUI Design skill
What this skill tells your AI
The instructions your AI receives, as published by caarlos0/dotfiles in skills/tui-design/SKILL.md and read by ahel’s review.
Design the command-line interface first. Add a full-screen TUI only when the task needs persistent context, fast navigation, selection, or live updates. A TUI owns the screen, the input, the scrollback, and the terminal state; do not take that cost without a reason.
Decide if a TUI is correct
-
Use a plain command when the operation is one-shot, batchable, or scripted.
-
Use the user's
$PAGERwhen the task is reading or searching long output. Page only when stdout is a terminal, and provide--no-pager. -
Use a TUI when users repeatedly navigate, compare, filter, select, or watch changing data. Process monitors, resource browsers, and Git interfaces are good examples.
-
Do not build a TUI only to get color or animation. Styled output, a prompt, or a single selector is usually enough.
-
Never make the TUI the only way to do an important operation. Ship equivalent non-interactive subcommands and structured output:
app # optional interactive interface app list --json app delete ID --yes -
Share the domain logic below both interfaces. The TUI is an adapter.
Keep the two interfaces separate
- Write results to stdout and write progress, warnings, and errors to stderr.
- Test each stream for a terminal independently. stdout can be a pipe while stderr is still a terminal.
- When output is not a terminal, change presentation only, never meaning. Remove color, cursor motion, progress animation, and the alternate screen.
- Provide
--jsonfor structured data and one JSON document per line for streaming events. Treat those schemas as a public interface. - Prompt only when stdin is a terminal. Provide a flag for every prompt so a script can supply the answer. Never block a pipeline on input.
- Support
-q/--quietand--verbose, with the same meaning everywhere.
Lay out for a character grid
- Make 80x24 the minimum target. Essential navigation, current state, primary actions, and errors must stay usable there.
- Recompute the layout from the current size on every resize. Handle
SIGWINCH, then query the real dimensions; do not cache the launch size. - Use explicit breakpoints. Drop decoration and secondary metadata before you hide primary content. Below the minimum, print a plain "terminal too small".
- Clamp sizes before subtracting. Test 1x1, 20x5, and zero-size reports.
- Give headers, status bars, and key hints fixed heights; give the main view the rest; set minimum sizes on interactive panes.
- Wrap prose. Truncate scan-oriented rows with an ellipsis and offer a detail view. An ellipsis signals hidden content; silent clipping does not.
- Use the alternate screen for sustained applications, and inline rendering for command-shaped work that should leave output in scrollback.
- Spend borders carefully: two bordered panes cost at least four columns at 80. Prefer whitespace, alignment, and one shared divider.
Handle text as the terminal sees it
- Measure rendered cells, not bytes, code points, or string length. CJK is usually two cells, combining marks are zero, and one emoji may be many code points.
- Cut and edit at grapheme-cluster boundaries, not code points.
- Do not treat East Asian Width as a complete width algorithm; ambiguous characters vary by terminal, locale, and font.
- Keep emoji out of borders, fixed columns, progress tracks, and cursors. Reported widths disagree across terminals and multiplexers.
- Treat Nerd Font and Powerline glyphs as optional. Keep an ASCII fallback and never let an unlabeled icon be the only meaning.
- Normalize tabs to a fixed width before measuring.
Use color as an enhancement
- Design a readable monochrome hierarchy first, then add semantic color.
- Support the ladder: no color, 16 colors, 256 colors, true color. Detect the
capability; do not assume true color, and remember
COLORTERMis lost through ssh and sudo. - Define roles such as muted, surface, selected, success, warning, and error. Map roles through themes; do not scatter literal colors in the code.
- Detect the terminal background before you choose fixed colors, and keep an unknown state for terminals that do not answer.
- Never encode meaning in color alone. Add a label, symbol, border, or cursor.
- Honor
NO_COLORwhen it is present and not empty. An explicit flag or the user config may still override it. - Target 4.5:1 contrast for ordinary text. You do not control the user's font size, theme, or rendering.
Bind keys that every terminal can send
- Support arrows and
hjkl. Show arrows in the footer and the vim aliases in the full help. - Make
Ctrl+Ca global, highest-priority exit that no focused widget can shadow. It must restore the terminal before it exits. - Bind
qto quit at the root,?to full help,/to filter,:to a command palette,Enterto activate,Tab/Shift+Tabto move focus. - Give
Escone consistent ladder: close a menu, cancel an edit or filter, close the overlay, go back one level. At the root it does nothing. - Leave
Ctrl+S,Ctrl+Q,Ctrl+D, andCtrl+Zalone. They belong to the terminal and the shell. - Do not require
Ctrl+Shift,Ctrl+Enter, or a Tab andCtrl+Idistinction. Legacy terminals cannot encode them. Use the Kitty keyboard protocol only for optional accelerators. - Keep mouse support off by default unless direct manipulation is central. Mouse reporting breaks ordinary text selection; document the Shift or Option override, and keep full keyboard equivalence.
Make the interface discoverable
- Keep a one-line footer with three to six actions for the current focus, for
example
up/down move Enter open / filter ? help q quit. - Derive the footer from the active keymap so it cannot drift from reality.
?opens a full help overlay grouped by global, navigation, current pane, actions, and search.- List aliases together, and never advertise a key that does nothing. Dim disabled actions and give the reason.
- Show the active sort, filter, mode, and match count. Do not overload plain typing silently.
- Treat empty states as content. Distinguish "nothing exists", "no match",
"loading", "failed", and "no access", and give the next action:
No matches for "prod". Press Esc to clear the filter.
Report state and errors honestly
- Print something within 100ms. Silence during slow work looks like a hang.
- Use a spinner only for unknown durations; switch to counts or a progress measure when totals are known. Always label what is happening.
- Stop animation on success, failure, cancellation, or non-interactive output, and keep the frame rate low enough for ssh and slow terminals.
- Show loading inside the affected pane and keep the surrounding context.
- Every error answers three questions: what failed, why, and what to do next. Rewrite low-level errors into domain language. Put the action last.
- Group repeated failures under one explanation instead of one paragraph each.
- Keep recoverable errors inside the TUI, preserve the user's input and selection, and offer the retry key. Restore the terminal before a fatal error prints.
- Confirm irreversible, broad, or externally visible actions, and name the
target:
Delete 12 pods in production?Default the focus to cancel. - Do not confirm actions with reliable undo. Offer undo instead, and only advertise it when recovery is guaranteed.
Structure the application
- Keep one source of truth for state, and one serialized path for every visible state transition.
- Make the update step a deterministic transition: no clock, no filesystem, no network, no globals inside it.
- Turn every occurrence into a typed event: keys, resize, ticks, completion, failure, cancellation.
- Return descriptions of effects instead of performing them in the update step. Keep rendering pure and cheap.
- Keep the event loop free. Anything slower than a few milliseconds is worker work that sends its result back as an event.
- Cancel stale work and also reject stale results with a generation counter; an old response must never overwrite a newer one.
- Tie worker lifetime to the screen or widget that started it.
- Keep the domain packages free of the TUI framework, of styling, and of
terminal APIs, so the same code serves
--json, tests, and scripts. - Reuse existing primitives for lists, tables, viewports, inputs, and help before writing your own; they already solved Unicode, scrolling, and paste.
- Never log to the stream the TUI owns. Log to a file and
tail -fit from a second terminal. Debug from a second process.
Own the terminal state
- Centralize acquisition and restoration. Track raw mode, alternate screen, hidden cursor, mouse capture, focus reporting, and bracketed paste, and undo them in reverse order.
- Restore on normal exit, on error, on panic, on
SIGINT, and onSIGTERM. Make cleanup idempotent and best effort. - Treat pasted text as data, not as keybindings, and bound its size.
- Accept that
SIGKILLcannot be handled. Documentresetorstty sane. - Render on state change, not in a free loop. Build the next frame, diff it, and write only the changed cells. Do not clear the whole screen every frame.
- Virtualize long lists, key the selection by item identity rather than row position, and bound streaming queues so producers apply backpressure.
Configure with a clear precedence
- Order: flags, environment variables, project config, user config, system config, built-in defaults. Provide a command that shows effective values and their source.
- Match the mechanism to the lifetime: flags vary per run, environment varies per machine, project files are shared, user config is personal.
- Follow XDG:
XDG_CONFIG_HOME,XDG_STATE_HOME,XDG_DATA_HOME,XDG_CACHE_HOME,XDG_RUNTIME_DIR. - Ship a complete default keymap and theme. High configurability without a good baseline is onboarding debt. Make reset easy.
- Map keys to semantic actions such as
item.open, not to functions, and detect conflicts.
Design for screen readers
- Assume a screen reader sees a text buffer, not your widgets. Borders and coordinates carry no meaning to it.
- Provide a first-class linear mode:
--plain,--no-interactive, or an accessible prompt mode that echoes context and asks one question at a time. - Do not assume a framework is accessible because it is popular. Verify with NVDA, JAWS, and VoiceOver separately; snapshot tests prove nothing here.
- Keep a real cursor where the user's attention is.
- There is no portable accessibility environment variable.
NO_COLORis the only established convention; anything else must be documented and paired with a flag.
Verify it
- Test state transitions without a terminal: send an event, assert the new state and the requested effect.
- Snapshot the rendered buffer at fixed sizes. Cover narrow, wide, empty, Unicode, selected, error, and loading states.
- Golden-test the plain contract with redirected streams,
TERM=dumb, andNO_COLOR=1. Assert that no escape sequence, prompt, or timestamp appears. - Add pseudo-terminal tests for lifecycle: raw mode, resize, paste,
Ctrl+C,SIGTERM, and panic. After each failure path, assert the terminal was restored. - Exercise the real thing through ssh and tmux, and on more than one emulator.
- Use recorded tapes for documentation, not as the regression gate.
Related skills
writing-testsfor deterministic tests, including the TUI cases above.code-reviewwhen reviewing someone else's terminal interface.go-conventionswhen the implementation is Go.
Signals
- GitHub stars
- 220
- Forks
- 11
- Last commit
- Sep 2026
Advanced
- Item type
- skill
- Key
tui-design-caarlos0- Source
- github.com/caarlos0/dotfiles