Use when designing or revising any TUI screen in daft — picking a layout, deciding what to surface, choosing a color, writing a label or error message, picking keys, designing a confirmation flow, or auditing an existing screen. Companion to daft-tui (which covers mechanics). This skill covers design decisions: hierarchy, scan order, color semantics, microcopy, keybinding doctrine, error and empty states. Synthesised from Tufte / Norman / Wathan-Schoger / Nielsen / Shneiderman plus a six-TUI reference study (gitui, lazygit, bottom, atuin, yazi, k9s).
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Use when designing or revising any TUI screen in daft — picking a layout, deciding what to surface, choosing a color, writing a label or error message, picking keys, designing a confirmation flow, or auditing an existing screen. Companion to daft-tui (which covers mechanics). This skill covers design decisions: hierarchy, scan order, color semantics, microcopy, keybinding doctrine, error and empty states. Synthesised from Tufte / Norman / Wathan-Schoger / Nielsen / Shneiderman plus a six-TUI reference study (gitui, lazygit, bottom, atuin, yazi, k9s).
daft TUI design
Design conventions and decisions for daft's TUI surface. Read this when you're choosing what to show, where, in what colour, and with what wording. The mechanics — panic hook, Stylize, hjkl wiring, terminal lifecycle — live in [[daft-tui]]; this skill assumes those are already settled.
The principles below are the body — each is concrete enough that two contributors should make the same call given it. The supporting research lives in references/design-principles.md (sources) and references/tui-references.md (canonical-TUI study).
Six principles
1. Strip non-data ink before adding any
Every glyph that doesn't carry information competes with the ones that do. TUI cells are scarce; decoration steals from data.
Apply: A worktree row reads feat/foo 3 ahead clean, not │ feat/foo │ 3↑ │ ✓ clean │. Use single dim rules to separate sections; never wrap rows in boxes. The sync table in src/output/tui/render.rs already follows this — it uses column headers (Branch Path Base Changes Remote Age Owner Commit) and no row borders. Keep it that way when adding columns.
Don't: add a column whose only payload is a static icon or [INFO] prefix. If the icon's information is "this row exists," delete it.
2. Build hierarchy from weight, color, and position — not size
Monospace forbids size shifts, so three levels of emphasis (primary / secondary / tertiary) carry the whole hierarchy: primary = .bold() + a named color or default fg; secondary = default fg; tertiary = .dim(). Pick one mechanism per level and stop.
Apply: A picker row's branch name is bold default-fg, the 3 ahead count is plain default-fg, the last-fetched-at is dim. The selected row gets the highlight color on top — but it stays bold default-fg underneath. Wiring: see [[daft-tui]] for Stylize-over-Style::default().
Don't: also .bold() the count or the timestamp because "they feel important." That flattens the three levels into two and you've lost the hierarchy.
3. Reserve each accent color for one meaning per screen
Color in a TUI is a typed enum, not a palette. With ~16 colours and no gradients, viewers learn each accent fast — and get confused fast if you reuse it. Publish a budget and enforce it.
daft's color budget:
Color
Reserved for
Anti-meaning
.red()
Errors and destructive markers (prune, fail)
Anything informational. Never "selected".
.green()
Success, completion checkmarks, "clean" state
"Selected" — green is outcome, not focus.
.yellow()
Warnings AND pending/in-progress (today: overloaded — see audit findings)
Don't add a third meaning.
.cyan()
Focus / selection / "active tab" (the canonical TUI default; daft uses it sparingly today)
Don't substitute another color for focus.
.dim() / .dark_gray()
Muted secondary metadata (paths, timestamps)
Never use dim as the primary text on a row.
A row may carry red OR yellow, never both. A row may carry green OR cyan (success vs. selected), never both. If you need to say "this errored AND is selected," the focus indication is the cyan background — leave the row text in its red/green/etc. semantic color.
4. Make every keystroke produce visible feedback within one frame
Silence after a keypress is a bug. Nielsen's "visibility of system status" in a TUI's budget is one redraw.
Apply:
j / k redraws the new selection before yielding. If a key handler triggers a slow operation (fetch, prune scan), swap the picker for <verb-ing>… on the frame before launching the work — don't leave the previous frame frozen.
Long-running phases (sync, prune) get the yellow spinner + phase label that presenter.rs already emits. Don't add a slow operation without a phase indicator.
Don't: assume an operation is fast enough to skip the spinner. The picker's responsiveness is graded against the worst-case input, not the median.
5. Show the keybindings; don't make users remember them
Norman's signifiers and Nielsen's "recognition over recall": the affordance must be visible. In a TUI, that's a hint line. Five of six reference TUIs (gitui, lazygit, bottom, yazi, k9s) render a bottom bar of currently-valid keys; atuin alone skips it.
daft's stance — open convention (see Open conventions): daft today shows neither a persistent hint bar nor a ? overlay in shared_picker. New pickers should add a dim bottom-line hint (j/k move h/l tab Enter select Esc cancel) until the global help model is settled. One line is cheaper than every contributor improvising.
Wiring for the hjkl + arrow match arms: see [[daft-tui]].
6. Make every action reversible, or warn before it isn't
Shneiderman's "permit easy reversal" and Nielsen's "user control and freedom." Esc is the universal back-out for takeover screens; Ctrl-C for inline commands. Destructive ops need an explicit confirm step, not a single keystroke.
Apply:
Esc always cancels a picker. Cancelled is a first-class variant on every *Decision enum (see RemoveDecision::Cancelled in shared_picker/remove_modal.rs).
Destructive operations go through a modal overlay — the convention used by remove_modal.rs, show_confirm_dialog, and the "Cancel sync?" / "Overwrite local changes?" flows in shared_picker/mod.rs. Match this pattern for new destructive ops; don't invent inline-chord confirmations.
Never bind a destructive action to a single unmodified letter adjacent to a navigation key.
Open conventions
These are choices daft has either made or left open. The questions came from divergences across the six reference TUIs; the answers here are grounded in daft's current code.
Layout: hybrid by purpose
daft renders TUIs in two modes, by purpose:
Inline viewport for sync / prune progress — src/output/tui/mod.rs, driver.rs, render.rs. The table stays in scrollback so the user can scroll back to a completed run.
Alternate-screen takeover for pickers — shared_picker/. Full-screen modal interactions that disappear on exit.
Decision rule: if the output should remain visible after the command exits, use inline. If it's a modal interaction the user dismisses, use alternate screen. Don't add a third paradigm (persistent multi-panel à la lazygit) without a strong reason — daft's flows are transactional, not session-oriented.
Tab dispatch: directional, not numeric
shared_picker/input.rs:48-49 dispatches tab walks with h / l (and Left / Right). The Tab / BackTab keys are bound to panel focus toggle, not tab cycling.
Don't repurpose Tab for tab walking. It's reserved for inter-panel focus.
When a picker grows to ≥4 tabs, revisit this convention — h / l becomes unwieldy past three.
Confirmation: modal overlay, never inline chord
daft's destructive confirmations are modal overlays (see shared_picker/remove_modal.rs, show_confirm_dialog). Five of six reference TUIs use modal confirms; bottom and k9s use inline chord (dd) for kill operations.
All new destructive ops use the modal pattern. Don't add a dd-style inline chord.
The modal's default focus is the safe option (Cancel / Materialize), not the destructive one. remove_modal.rs already does this.
"Cancel sync?", "Overwrite local changes?", "Continue?"
Empty state
Sentence-case statement
"No matching files found", "No shared files remaining"
Success (transient)
Past-tense terse
"Collected {path} from {wt}", "Removed {path}"
Error
"Failed to <verb> <thing>: {detail}"
"Failed to create shared storage: {e}", "Collect failed: {e}"
Inline instructional
Lowercase parenthetical
"Search: (type to filter)"
Column header
Title Case noun
"Branch", "Path", "Changes", "Remote"
Use these forms. The rule of thumb: titles announce, prompts ask, success states report, errors explain. Don't drift into mixed forms ("Searching..." with the ellipsis as a status when "Search:" is the prompt).
Theming: hardcoded slots
daft has no user theming infrastructure today and isn't planning any. All color choices live in code via the budget above. This is atuin's stance, not bottom's — and it's the right one for daft's surface area. If user theming ever lands, the budget above becomes the slot names.
What did NOT transfer from the graphical-design canon
These are tempting borrowings to suppress:
Tufte's sparklines and small multiples survive only for numeric time-series. daft has none — the sync table is categorical, not quantitative. Don't reach for them.
Refactoring UI's font-size hierarchy, shadows, saturation gradients have no monospace analogue. .bold() + .dim() + the color enum are the entire emphasis toolkit. Don't simulate sizing with banners or ASCII art.
Norman's physical affordances reduce to "show the key in the footer." Don't write text that pretends to be a button.
Aesthetic minimalism as a separate principle is redundant with Tufte's data-ink ratio once you're counting cells.
"Design dialogs to yield closure" matters in form-heavy GUIs. daft's flows are one-shot commands; closure is just process exit.
Reference TUIs — what to study, where
When a design question doesn't resolve from the principles, study these source files in the linked repos. Each is the most direct answer to a question daft also has.
Question
App
Where to look
Color-promoting context-relevant keys
lazygit
pkg/gui/options_map.go, renderContextOptionsMap
Semantic color slot naming
gitui
src/ui/style.rs
Per-binding desc field (descriptions next to bindings)
Boxing rows. Don't draw borders around table rows or list items. Use dim separators between sections at most. (Violates Principle 1.)
Color as decoration. Don't add green or cyan to a label "to make it pop." Every accent must answer "what state is this signalling?" If the answer is "none," it stays default fg. (Violates Principle 3.)
Yellow for two things on one screen. Yellow currently signals both "in-progress phase" and "warning" in daft — when both can appear on the same screen (e.g., a warning during a running sync), one of them must move. (Violates Principle 3.)
Inline-chord destructive ops. No dd-style two-key destructive confirms. Use the modal pattern.
Footer-less picker. Adding a new picker variant without at least a one-line dim keybinding hint, until the global help model is settled. (Violates Principle 5.)
Re-documenting mechanics. If you find yourself writing about how a panic hook works or how Stylize composes, you're in the wrong skill — that's [[daft-tui]].
