⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

version unknown

ui-polish

agentes-chat@agentes-chat⭐ 1.3k stars· last commit 2d ago· 52 open issues

agentes.chat UI rules and routine. Use before any UI, CSS, layout, spacing, alignment, component, screen, page, dialog, card, chat, menu, toolbar, form, email or on-screen copy change in agentes.chat (apps/web, packages/ui, apps/mobile, apps/desktop), and for any design review or polish pass. Every rule is a number. The routine fixes the kit component, sweeps every sibling, keeps the CI geometry check green, shoots before and after at 1280 and 390, and runs break-ui and Impeccable's audit.

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐⭐⭐⭐⭐ 1.3k on GitHubGitHub ↗

Very popular

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README

Agents join your channels and DMs with a role, an owner and a budget, work where everyone can see, and ask before anything sensitive. They think with Claude, ChatGPT and Codex, Gemini, Grok or any OpenRouter model, use MCP tools, and get a computer and a browser of their own. Self-hosting is free and always will be (AGPL-3.0, no license key, no paid tier), or use agentes.chat Cloud, the paid, hosted option: both run this same codebase.

Why agentes.chat

  • One conversation for people and agents. Agents are members of the workspace, not a sidebar bot: they read the channel, answer when it's theirs to answer, and hand work back and forth.
  • Safe by default. Merges, deploys and anything a connector marks sensitive wait for a person's approval in the chat. Every model call is budgeted and recorded per agent.
  • Agents that actually do the work. Their own computer, a real browser you can watch and take over, Claude Code and Codex, routines on a schedule or an event, memory that learns.
  • Yours to run. One command on a Linux server, Postgres as the only dependency, web, desktop and phone apps that connect to your server.

An open-source alternative to team chat with bolted-on AI bots, and to single-player agent apps such as Grok Bot, ChatGPT dots, OpenBot.run and OpenMausBot (all of which it imports from).

Quick start

  • Cloud: sign up at agentes.chat.

  • Your own server (Linux):

    curl -fsSL https://agentes.chat/install | sh
    

    It installs what's missing (Docker, after asking), and puts agentes.chat on your Tailscale network, private and with HTTPS, in about 10 minutes. SELF_HOSTING.md covers the phone, a public domain instead, upgrades and backups, and running it without Docker.

Each organization adds its own model key (Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax or OpenRouter) in the app.

Features

FeatureStatusNotes
Channels, threads and DMs✅Public and private channels, DMs with people and agents, and threads started from a message or a stretch of conversation, then followed, closed and archived.
Agents✅Each has a role, instructions, a model (Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, or any of OpenRouter's), a monthly budget checked before every call, and an owner who gets its approvals and budget alerts.
Router✅A message nobody @mentioned goes to the router. Code picks the candidates, and Jev, a small model on the workspace's key, the author's own plan (on a server you run yourself) or simple rules pick among them. Up to four agents answer, or none, and the message says who picked. It never grants a permission.
Approvals✅Merge, deploy, rollback and anything a connector marks sensitive wait for the agent's owner or an admin, as a card in the chat. The approved action runs once, exactly as approved.
Connectors (MCP)✅The organization connects an MCP server once (official servers such as Linear, Sentry, Notion and Stripe in one click, or any server by URL), and each agent gets only the scopes granted to it. With a Composio key, Google Drive, Calendar, Sheets and Docs, Slack, Microsoft Teams, OneDrive, HubSpot, Salesforce, Zendesk, Asana, Figma, Intercom, Dropbox, Square, Instagram, Facebook, LinkedIn, YouTube, Pinterest, Reddit and WhatsApp Business connect through Composio.
Skills✅Markdown skills (SKILL.md) from a zip, a GitHub link or the editor, with versions. Agents load one when a task needs it.
Routines✅An agent works on a schedule ("every weekday at 9am") or on an event (a GitHub pull request, a Sentry issue, a new task, an incoming email) and delivers to a channel, a DM or the Drive.
Memory✅Agents keep decisions, facts, preferences and lessons from conversations, learn from corrections and refused approvals, and propose changes to their own instructions for a person to accept.
Wiki✅Pages with full history, [[links]] from messages and tasks, backlinks and a graph, and Obsidian import. Agents read it, and write only when a person asks.
Drive✅Files and documents with versions, kept on disk, in Postgres or in an S3-compatible bucket (or the organization's own). Documents export to Word and PDF.
Tasks✅A board with your own columns. A task given to an agent moves on its own as the agent starts and opens a pull request, with its current step in view.
Notifications✅One inbox for mentions, replies, approvals and agents' questions, grouped, with push on the phone app.
Agent computers✅Agents run commands and code on a computer that runs the runner: the server itself (the setup makes it the agents' computer by default), or a server, VM or laptop you connect. On agentes.chat Cloud, each workspace gets a computer of its own.
Agent browserBetaOn connected computers, agents open sites, fill in forms and read pages through agent-browser. Watch it live in the chat, on the agent's page or in the phone app, take over for a login or a captcha (or to open a site yourself), and hand it back with a note.
Claude Code and Codex✅On a connected computer, an agent can think with Claude Code or Codex, on the plan signed in there, instead of through the API. Codex is still in testing.
Postgres connectorBetaAgents query your database with SQL, read-only: one parsed statement in a read-only transaction, with time and row limits.
Agent mailboxesBetaAgents read, draft and send from mailboxes the organization connects: Gmail and Outlook through Composio, any other over IMAP/SMTP. Sending asks for approval by default.
Phone appBetaiPhone and Android (Expo), on agentes.chat Cloud or your own instance: scan the QR code from "Open on your phone" in the web app, which signs you in, or change the server on the sign-in screen.
Desktop app (macOS; Windows and Linux on x64 and ARM64)Betaagentes.chat Cloud in a window of its own, with nothing running on your computer; or your own server: connect to it (by address, by what the app finds, or by the link the setup prints) or run agentes.chat on this computer with no Docker (Mac and Linux). On your own server, its agents can also work on that Mac or Linux computer, after the app asks. Download it from /download on any server; while it's in beta the Mac build isn't notarized and the Windows i

// HOW IT'S BUILT

// FILES

2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • .claude/skills/ui-polish/SKILL.md11 KB1ca045e443d2upstream ↗
    ---
    name: ui-polish
    description: agentes.chat UI rules and routine. Use before any UI, CSS, layout, spacing, alignment, component, screen, page, dialog, card, chat, menu, toolbar, form, email or on-screen copy change in agentes.chat (apps/web, packages/ui, apps/mobile, apps/desktop), and for any design review or polish pass. Every rule is a number. The routine fixes the kit component, sweeps every sibling, keeps the CI geometry check green, shoots before and after at 1280 and 390, and runs break-ui and Impeccable's audit.
    ---
    
    # UI polish for agentes.chat
    
    Measure, don't eyeball. Vision misses 1 to 2 px errors, source review can't see pixels, and a screenshot diff shows what changed, not what's wrong. Check the rendered screen.
    
    ## The kit's numbers (packages/ui tokens)
    
    - Three control heights: 32 (toolbars, list filters, in-row actions, composer), 40 (forms, page and dialog buttons, search, selects, sidebar items), 48 (only the main action of a full screen). Radius is 30% of the height. Pills only for labels you don't press.
    - Spacing steps 4, 8, 12, 16, 24, 32. Form fields 16 apart, a label 4 from its control. List rows 8 above and below (`--row-py`, 12 on touch): one line 36, with a 24 face 40, with a 32 control 48. Cards 16 inside (`--card-pad`). A plated button stays 8 from its row's or card's edges (`--action-clear`). A row of actions (`.ac-acts`) puts 16 between what the eye sees, on touch too, and two plated buttons 8 apart. A small button keeps its 32 plate on touch: its target grows to 44 past it (`--touch-target`), up and down only when it has words.
    - Every page starts at `--page-top` (its title's line on the sidebar's workspace row, 20 on phones) and `--page-x` (24, 16 under 640). Header, toolbar and content 16 apart, list groups 24 (`.ac-groups`).
    - No raw px for padding, gaps or heights. Use the tokens.
    - Rounded inside rounded: inner radius is the outer radius minus the gap between them.
    - Fonts are fixed: Source Sans 3 (the app, titles too), JetBrains Mono (code), Funnel Display (brand), Doto (dot type), Onest and Martian Mono (the kit's defaults, the desktop app). Never swap or "improve" them, whatever a tool suggests. A new font gets `--cap-mid` and `--cap-top` measured again.
    
    ## Rules
    
    Layout
    1. One spacing, padding and toolbar on every page: search on the left, tabs, filters and selects on the right. The page title starts at the same top and side on every page. From 761 the toolbar keeps one row, the search giving way down to 160. When even that can't fit it takes two rows (the search and the actions, then the filters in one sideways row). The search never goes under 160, on a phone either: when the actions don't fit beside it, it takes the top row alone and the actions lead the filters' row. Never a control alone on a row: a lone small filter shares a row, a lone segmented control fills its own.
    2. Dense but comfortable, like Resend. When a screen feels crowded, regroup and space it. Don't just delete.
    3. Content takes all the width it has. Fields in one form fill its width, each control as wide as its stack.
    4. A leading square (face, icon tile, logo) never owns a column while the text beside it is squeezed: face, name and role on one line, text and controls full width below. A square beside 3 or more lines is a bug. This covers cards and the inbox. A DM has no face column at all, its header already says who it is.
    5. Chat messages: the name's capitals start at the face's top (`.ac-faceline` with `.ac-captop`), and message text uses the full column width.
    6. A dialog's inner card reads square-ish, never a wide flat strip.
    
    Alignment and spacing
    7. Controls side by side share one height. Shrink the taller one.
    8. Gaps between buttons are even and never zero: plated buttons 8 apart, icon buttons' plates 8 apart, header glyph boxes 32 and 8 apart. Header icon buttons get no plate on hover or press, only the icon colour changes.
    9. Faces, tiles and controls stay at least 8 from every row edge and divider, also inside a highlighted or bordered row. Top padding equals bottom, left equals right.
    10. Icons, ⓘ and count badges centre on the capitals' middle within 1 px (`.ac-iconline`, `0.5lh - var(--cap-mid)`), also inside a line of words. A face beside one line centres on it. Words right after an inline face share its baseline. Initials sit optically centred in their tile.
    11. A picture taller than 1.25 lines beside 2 or more lines: the first line's capitals start at its top (`.ac-captop`). A step badge beside a name and a subtitle centres on the whole two-line block.
    12. Menu rows are their kind's height, 2 apart, the menu's padding 4: a line 32, a name with its description 50, a row with a face 36 (40, 58 and 44 on touch). Every option in one select is the same height, a description line never makes one taller. A select never shows empty. A people select shows a small face.
    13. List times end on one right edge. A list heading (GroupHeading) sits on its list's outer edge, 8 above it, 16 under the toolbar.
    14. Text and placeholders sit vertically centred in their inputs.
    15. Borders and shadows survive rounded corners. Hover backgrounds stay inside their row. A scrolling parent never cuts a control's ring.
    16. One class keeps one height, padding and gap on every screen.
    
    Text
    17. Wrap meaning, never cut it with an ellipsis. Give a fact value the room to stay on one line. No row wraps word by word. Words and their actions (a title and its buttons, a notice and its fix, a toast and Desfazer) are an action line (`.ac-actline`, phone `ACT_LINE`): the actions sit beside the words only while the words fit on one line there, else on a row of their own under the words, from their edge. A close × stays beside the title, outside the line. Ver detalhes and every toggle like it go under their words. Only one small control that keeps its place (a switch, a select, a chevron) ends a row with `.ac-actrow`.
    18. No button covers text. Desfazer never takes its own column: it goes at the end of its sentence, or under it when there's no room.
    19. Badge, pill and button text 4.5:1 (3:1 for large text) in both themes. An unread or waiting count at least 3:1 against what's behind it.
    20. Less text: a title, at most one short sentence, the controls. Help goes behind an ⓘ beside its label, a title plus one sentence (Field's and SettingRow's `hint`). Under a field only a format example (`example`) or an error. Under a setting only what must stay in sight (`status`). Background reading goes to a site page, linked once. Cut sentences that only reassure.
    21. No semicolons in product copy, in any language or surface. No em dashes, except member group headers like "DONO — 1". Long "·" chains read dense.
    22. Links use the house link style, never the browser's blue underline. Third-party products and AI providers show their official logo, never a generic icon. Text between backticks is an inline code pill (`.ac-inlinecode`) wherever formatted words show: a message, a document, an agent's words on a card.
    
    Components
    23. On phones, a card's lone primary button is full width (at least 90% of the card).
    24. Bulk actions (Aprovar tudo, Agora não para todos) are compact and aligned, never tall floating buttons.
    25. Cards inside a message share one width and stand apart from the row's hover in both themes. A group of one shows the item itself, not "Used 1 tool".
    26. A link that only shows one dialog opens it in place. It never navigates away.
    27. Calendar events follow Google Calendar's look, with no curved accent border.
    28. A long-running screen always has a way out (stop or back). A failure says what happened and offers a retry.
    29. Motion is smooth. Nothing pops in, not a new card, not an opened thread.
    30. Scrolling columns hide their thumb until hover.
    31. Emails carry the agentes.chat look in light and dark clients.
    
    Discord-style layout
    32. Members column on the right, open by default on desktop, toggled from the header. Roles have configurable colours and their marks are full circles.
    33. Agents hold no roles. Their card shows "Quem pode falar" as plain lines (mark or face, then name) with an "Editar" link, never pills.
    34. Unread is a bold name and a left pill. A red count only for mentions.
    35. The left sidebar has one "Mensagens diretas" section. The pinned bar under the header stays.
    
    ## Routine for every UI change
    
    1. A big batch (many screens or reports): show proposals first, with the before, the after and the reporter's exact words per item. Build after the go-ahead.
    2. Fix the kit component in packages/ui, where every caller routes through, not the one screen.
    3. Sweep: grep every place the pattern appears in apps/web, packages/ui, apps/mobile and apps/desktop, including places nobody reported, and fix them in the same change. The phone gets the same fix.
    4. Run break-ui on worst-case data at real widths (below).
    5. Run Impeccable's audit and triage the hook's findings (below).
    6. Keep the geometry check green. CI's Geometry step lays out every screen the tests reach at 1280 and 320 (`scripts/geometry`). Locally run only `bun test scripts/geometry/check.test.ts`. A rule here with no check yet gets one in `scripts/geometry/rules.js`, with a bad and a good fixture. Update `baseline.json` only after fixing offenders or for a verified false positive.
    7. Measure what ships. The build can order CSS differently from dev (a `translate` beside a `transform` was dropped once). Check the build's CSS too.
    8. Shots before and after at 1280 and 390, light and dark when colour changed, with toggled states open (menus, Advanced sections, expanded rows, hover). After shipping, publish one before and after page. Each release also gets a visual sweep for problems nobody reported.
    9. Done means every clause of the ask is checked on a real screen.
    
    ## break-ui
    
    Run the break-ui skill on any component or screen that renders data: names, emails, counts, titles, translations, empty and huge lists. Look at the component's real container width, at 320 and at the widest, and in dark mode. Feed the worst case through the component's test fixture, so CI's geometry check keeps measuring it. Its dev-only toggle never ships.
    
    ## Impeccable
    
    - Its hook runs by itself after every Edit or Write of a UI file, and again at Stop. Fix real findings. For a false positive, say why in your reply. Add an ignore only with `--local` (git-excluded), never in the shared `.impeccable/config.json` of this public repo without approval.
    - Before calling a screen done, run `/impeccable audit <target>`. It only reports.
    - For a fast deterministic pass over many files, run the launcher in the impeccable skill's folder: `scripts/impeccable detect <paths>`.
    - In this repo, don't run the commands that write project files or change the product's look: `init`, `document`, `extract`, `critique` (it saves reports into the repo), `live`, `generate`, `craft`, `shape`, `bolder`, `overdrive`, `colorize`, `delight`, `pin`, `hooks on` or `hooks off`. Its live mode stays off (`IMPECCABLE_LIVE_COPY_AGENT=off`). Never take its font advice.
    - Keep frontend-design and design-taste-frontend off app screens.
    
  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/agentes-chat-agentes-chat-claude-skills-ui-polish.md

// REPO STATS

1.3k stars
52 open issues
Last commit: 2d ago