English | 简体中文
Show me helps coding agents put useful views inside the conversation, then continue the task from what you edit, choose or adjust.
A short question still gets a short answer. When a view helps, it starts with the content you are already discussing. Edit a passage in place, choose between real previews, or move a control and see the result change. Keep talking whenever you prefer; the agent carries the context forward.
Open the examples. Edit the wording, choose a study-plan view, or adjust a session's duration. Each example has a local receiver that checks your input and updates the sample. The page makes no model calls and contains synthetic data.
Inside a supported chat host, an explicit action sends your changes to the current conversation. The agent checks the task and source version, does the agreed work, and reports what happened. Typing and exploring stay local.
Codex and ZCode show the editor in the conversation. In DeepSeek Harness, a chart arrives as an HTML attachment. Click the file card to see it in the right sidebar. For tasks that need your exact edits back, the host guide covers the separate local editor for Claude Desktop's Code tab and the experimental DeepSeek route.
After installing, ask your agent to use show-me on the current task.
| Ask | Useful response |
|---|---|
| “How many minutes are in two hours?” | A direct answer |
| “Compare these two plans by price, limits and cancellation.” | A comparison using the same criteria |
| “This paragraph is close. Let me fix a few words.” | The actual passage, editable in the conversation |
| “Show me both layouts before I choose.” | Comparable previews with a visible choice |
| “What changes if I only have an hour?” | A control beside the result it changes |
| “Use my changes and finish the file.” | The agent resumes the task and reports the actual result |
| “Save this explanation for my team.” | The requested file |
The skill guides the agent's choices. The small optional handoff helper validates text edits, choices and numbers; it leaves the layout and execution to the task. See conversation handoffs for the snapshot format and receiving workflow.
Download this repository or clone it:
git clone https://github.com/ninondev/show-me.gitCopy the entire skills/show-me/ folder into a skill directory your app reads,
including references/, scripts/, assets/, agents/ and LICENSE.
- Codex:
~/.agents/skills/ - Claude Code:
~/.claude/skills/ - ZCode and DeepSeek Harness: follow the
app-specific installation and capability notes.
For ZIP import, use the release package with
SKILL.mdat its root.
If show-me is already installed, compare or back it up before replacing it.
Start a new session and ask the agent to load this copy. The optional scripts
need Node 20 or later; the presentation guidance does not require Node.
The skill uses the active host's renderer, question controls and follow-up API. Read the host reference for the route your app supports. In a client without interactive views or a waiting local command, the conversation continues with text, tables or native questions.
Finished files open in the host's preview when supported, otherwise in the operating system's default application. Local HTML uses the default browser. An explicit request for a file, report, website or export sets the deliverable.
The batch text example learns bounded transformations from examples and previews their effect on a whole batch. Save an operation to reuse it without model calls. This is an optional workflow for repeated text tasks; its guide covers JSON input, supported transformations and CLI execution.
node --test tests/*.test.mjs
node scripts/build-demo.mjsTests cover handoff validation, uncertain delivery, repeated submission, local receivers, draft persistence, host isolation and the batch-operation engine. The evaluation cases and rubric cover the agent's behavior. Exercise host integration in a real conversation and distinguish sending a response from completing work.
To compare effort or cost, include preparing the view, reviewing it and continuing the task. Total human time and model-token savings have not been measured.
Derived from Charlie947's show-me,
at commit 69e488e.
The original centers on HTML previews. This derivative draws on GPT-6's
Intelligent UI to choose a presentation for each answer, with conversation
views and interaction supplied by the app.
It retains the original's practice of showing visual choices before asking, making alternatives structurally distinct, and checking that finished work was presented.
MIT licensed. The original copyright notice is preserved in LICENSE and in the installable skill folder.