# Desktop UI AI Workspaces

> AI Workspaces (Desktop)

- Skill: `simply-ehis/desktop-ui-ai-workspaces` (Agent Skill)
- Install (CLI): `npx skillmds@latest add simply-ehis/desktop-ui-ai-workspaces`
- Raw SKILL.md: https://api.skillmd.com/api/skills/simply-ehis/desktop-ui-ai-workspaces/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: simply-ehis (https://skillmd.com/u/simply-ehis)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/simply-ehis/desktop-ui-ai-workspaces

---


# AI Workspaces (Desktop)

AI features should help a person complete a concrete task. Do not make a chat panel the default answer if a structured action, preview, or inline suggestion better fits the work.

## Establish the interaction contract

Before choosing components, state the user goal, what the system may do, what it needs from the user, what can change external state, and how a result can be inspected, revised, or undone. Match the interface to that contract:

- Use **inline assist** for a local, inspectable suggestion in the current work.
- Use a **task/run view** for work that takes time, spans files or tools, or needs a durable history.
- Use a **chat surface** when conversational refinement is genuinely the fastest way to express the task.
- Use a **form or command** when inputs are known and repeatable; it is more scannable and safer than a prompt box.

## Make AI state observable

Never represent a run with a vague spinner and a comforting sentence. Show the meaningful state: queued, preparing, waiting for approval, working, needs input, completed, failed, or cancelled. For a long task, show a concise activity summary and elapsed time; reveal detailed logs on demand without forcing every user to read them.

Separate three things visually: the user's instruction, evidence or source context, and the system's proposal or action. Preserve provenance where it matters: generated text should identify its sources or uncertainty; file or data mutations should name the target and show a diff, preview, or reversible consequence.

## Approval and recovery

- Put approval immediately before the consequential action, with a plain-language summary of scope. Do not bury it in the initial prompt.
- Make the safe path obvious: cancel, edit inputs, retry, or keep the partial result. Do not make a failed run feel like lost work.
- Treat destructive, financial, external-message, credential, or broad file changes as a higher-risk tier. Require an explicit review surface, not a generic "continue" button.
- Do not imply certainty the model does not have. Say what was found, inferred, or unable to be checked in a compact, useful way.

## Density and calm

The primary work remains primary. Avoid a permanent, oversized prompt composer, animated thinking theater, token counters, rainbow status pills, or a card around every message. Let a compact composer expand only when the task needs more context. Use color for run state and warnings, not as decoration.

For model choices, lead with capability and trade-off in task language (for example, "fast draft" versus "deep analysis"); expose identifiers, quotas, and advanced parameters behind a details affordance unless the audience is expert.

## Accessibility

Announce state changes without repeatedly interrupting screen readers. Keep streamed content keyboard-reachable, preserve focus when a run updates, and ensure cancellation, approval, and result actions are available without pointer precision. Respect reduced motion; streaming should never make text jump beneath a reading user.

## Review format

| Before | After | Why |
|---|---|---|
| A generic chat window is used to configure a repeatable export task | A structured export panel with an optional "refine with AI" field | Known inputs are faster to scan, validate, repeat, and audit than an open-ended prompt |
| A long agent run says only "Thinking…" | A compact run status shows current phase, elapsed time, controls, and expandable activity | The user can decide whether to wait, intervene, or safely leave the task running |
| "Approve" is shown without consequences | The approval names the target, proposed change, and review/undo path | Consent is meaningful only when the scope is legible |