Intelligent UI Anything

Turn what your agent just produced into one small page you can sort, read or pick from, and get your pick back into the session without retyping it.

Claude Code · Codex CLI · Cursor · Antigravity · claude.ai

Quick start · Examples · Five shapes · How it works · Install · Cost

Why

Agents answer in prose. When the answer is ten models with six attributes each, or twenty findings across three sessions, prose makes you scroll, compare in your head and then type your choice back. This skill gives the agent five locked page templates instead. The agent writes JSON, a script fills the template, checks it, and opens it where you are. You sort, filter and pick. Your pick comes back to the agent as one line, by a Copy button or straight into the session.

• Five shapes, nothing else. Compare, choose, explain, track, inspect. A question that fits none gets prose.
• Looks like an answer, not a report. Cards, the leaders at a glance, tappable rows and options, status icons, a floating pick dock, and motion that streams in. Light, dark and phone layouts are designed, not shrunk.
• The decision comes back. Every pick shows the exact line the agent will get. Served pages send it with one click.
• Opens where you are. An artifact in Claude Code, your browser in a terminal, a localhost URL over SSH, a path inside a sandbox.
• Honest by construction. A page that states a number its data does not hold fails the check and never opens.
• Zero dependencies. Node 18 built-ins, no network, one self-contained HTML file per page, light and dark.

Quick start

npx skills add IndusCore-AI/intelligent-ui-anything

Then ask the way you would ask a person:

compare these models and let me pick one
lay out the three options for the migration
show me what changed, what's risky and what's undecided across today's sessions
keep a checklist page updated while you run the release

Try it without an agent: node scripts/render examples/comfyui-models.json opens the example page.

Examples

Every example is a runnable JSON input in examples/ with real, sourced data.

Five shapes

The question	Shape	What you get
Which of these is best on what matters?	compare	sortable, filterable table; optional bar chart; optional pick per row
Which one should we do?	choose	2 to 6 option cards with good and watch-out points; one pick
What does this say, and what changed?	explain	a long document with contents, anchors and a changed-sections view
How far along is it?	track	a checklist with progress that refreshes while the agent works
What happened, what is risky, what is open?	inspect	a filterable timeline by lane, severity and session; diffs; decisions

Colours carry meaning and nothing else: five elements, severity, pass and fail, diff, and finance colours by locale (red is up in China, Taiwan and Japan; green is up in the US). The palettes and their contrast measurements are in references/semantics.md.

How it works

you ask  ──▶  agent picks a shape and writes JSON  ──▶  scripts/render
                                                          │ fills a locked template
                                                          │ checks it (10 floor rules, 2 security rules, Hallmark lint)
                                                          ▼
                    artifact · browser · localhost URL · printed path
                                                          │
you pick  ──▶  "Use model FLUX.2 [klein] 4B"  ──▶  back to the agent (Copy, or Send when served)

1. The agent writes content, never markup. The template, styles and script are fixed, so every page looks the same, costs less to write and passes the same checks.
2. The check runs before anything opens. Tokens only, focus rings, contrast in both themes, phone width, no horizontal scroll, readable without script, no em dashes, no buzzwords, no invented numbers, nothing loaded from the network, no credit or tracking in the page. A failing page is written but not opened, and the agent is told what to fix.
3. The script reads the environment and picks where the page goes:

You are in	The page
Claude Code with the Artifact tool	becomes an artifact
Claude Code on an API key, Bedrock or Vertex, or claude -p	opens in your browser
Codex CLI (sandboxed)	is written, and the path is printed
An SSH session	is served on 127.0.0.1, with the ssh -L line printed
A Linux box with no display	is written, and the path is printed

What was measured in each host, and what is still unverified, is in references/hosts.md.

Install

npx skills add IndusCore-AI/intelligent-ui-anything

/plugin marketplace add IndusCore-AI/intelligent-ui-anything
/plugin install intelligent-ui-anything@intelligent-ui-anything

git clone https://github.com/IndusCore-AI/intelligent-ui-anything ~/.claude/skills/intelligent-ui-anything

git clone https://github.com/IndusCore-AI/intelligent-ui-anything ~/.agents/skills/intelligent-ui-anything

Run Codex with --sandbox workspace-write (its interactive default). codex exec defaults to read-only, where no page can be written.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗