VisActor Agent Skills

English | 简体中文

VChart, VTable and data storytelling skills for agents that can read files and run commands. Generate charts, tables, HTML data stories and Web-PPT presentations, and diagnose chart configuration problems.

Skills

Each skill includes its instructions, local references, templates and scripts. No separate VisActor CLI, MCP server, model key or hosted VisActor API is required. The host agent supplies the model.

Recommended: use the Skills directly in Codex or Claude Code. Your agent provides model access and tool execution. Other compatible agents may also use these Skills, but do not have dedicated setup instructions here.

Install

Run these commands in the project where you want to use the Skills:

npx skills add VisActor/skills --skill vchart-skill
npx skills add VisActor/skills --skill vtable-skill
npx skills add VisActor/skills --skill data-storytelling

Follow the installer’s prompts to select Codex or Claude Code. You can install each Skill independently. More installation options are documented in the Skills CLI reference.

Requirements:

• Node.js and npm to run the installation commands above.
• An agent with filesystem and shell tools, plus its own model access.
• Python 3.9+ for VChart/VTable; Python 3.10+ and Node.js (JavaScript validation) for data-storytelling. Python scripts use the standard library.
• A browser and internet access for generated HTML. Templates fetch pinned VChart/VTable, React (when needed), and Monaco packages from the public unpkg CDN. Data-storytelling also loads pinned VSeed and GSAP. This is not a fully offline bundle.

VChart knowledge search reads bundled example/SOP files locally. It uses keyword matching, not a hosted semantic search service; unmatched questions fall back to the included type and API references.

Use

After installation, ask your agent to use the selected Skill with your data. For example:

• “Use VChart to compare sales by region and quarter in this CSV. Give me the editable source and HTML.”
• “Use VTable to build a sortable detail table using every row in this file.”
• “Use data-storytelling to explain retention changes in this CSV to our growth lead as a shareable HTML story.”

Data-storytelling uses the current agent to author VSeed configurations and local scripts to assemble the output; it needs no hosted chart generation service. Images use host capabilities or the declared fallback. Web-PPT is a browser presentation, not a .pptx file.

Generated JavaScript/HTML is executable code. Preview your own or reviewed artifacts locally; the evaluation browser is not a sandbox for arbitrary untrusted submissions.

Development and maintenance

This section is for contributors editing Skills or running local tests. The default and primary maintenance branch is develop. Start with CONTRIBUTING.md.

Install from a local checkout

From the root of your local repository, install the working copy of a Skill:

npx skills add ./skills/vchart-skill
npx skills add ./skills/vtable-skill
npx skills add ./skills/data-storytelling

Alternatively, copy the complete Skill directory into your agent’s skill directory, including its references, templates and scripts. Do not copy only SKILL.md. See the local debugging guide for Codex and Claude Code workspace setup. Refresh the installed copy after changing a Skill before retesting.

Regression tests

The Python maintenance environment requires Python 3.12+ and uv:

uv sync --locked
uv run pytest tests -q
uv run ruff check skills tests evals/scripts

Fixed browser smoke cases are maintained by Skill in evals/smoke/, with each case’s configuration, inputs and acceptance criteria together. Run all with uv run python evals/scripts/smoke.py, or select --skill data-storytelling / --case story-ppt after the browser setup in that guide.

Model API task debugging

For users with model API access, a lightweight harness can load a selected Skill and run a fixed case or a custom prompt with input files. It supports OpenAI-compatible Chat Completions. Treat it as an experimental route for bounded VChart/VTable tasks with clear inputs and outputs, not as a validated replacement for a coding agent. For end-to-end data storytelling and everyday interactive debugging, use Codex or Claude Code.

For fixed prompt/data runs, observable execution logs, browser checks and screenshots, see evals/README.md. These are maintainer tools, not part of the installed skills. The default runner uses a configurable model API with your own key; Codex/Claude Code are not required. Model calls may incur usage charges. Copy evals/model.example.json to the ignored evals/model.local.json, configure the endpoint/model/key environment variable, then run uv run python evals/scripts/run.py --case vchart-grouped --browser. See the quickstart for setup and usage limits.

The case catalog includes 32 VChart tasks, 4 VTable tasks and 5 data storytelling tasks, organized by Skill with explicit acceptance criteria. The runner defaults to the 8-task VChart core group; curated coverage does not mean every task has been executed.

Versioning

Skills are versioned together at repository level. Each formal Skills release is identified by a Git tag covering the Skill instructions, references, templates and scripts. Individual Skills do not have separate version numbers.

Use a release tag or commit SHA to identify the version when reporting an issue or reproducing a result; branch contents can change. Dependency versions such as VTable 1.26.8 identify the underlying library, not the Skill release. Historical VMind SDK tags remain unchanged and belong to the earlier SDK.

Earlier VMind versions

The previous VMind SDK is no longer maintained. Its code and documentation remain in historical releases.

MIT license