The agent logged into X.com via Actionbook browser, queried from:elonmusk since:2026-02-09, scrolled through all results, and produced a comprehensive activity report:

What the agent did:

Queried Actionbook API for X.com selectors
Opened X.com Advanced Search via actionbook browser
Scrolled to exhaustion to capture all tweets
Counted 38+ original tweets/replies + 2 retweets
Identified 3 distinct posting bursts and 6 topic categories
Generated bilingual (EN/ZH) json-ui report

Report highlights:

Metric	Value
Total Posts (as of 12:30 UTC)	38+ tweets/replies
Peak Rate	~1 tweet per 3 min
Most Viewed	"Grok" (Tim Pool quote) — 21M views
Top Topics	Epstein/Bannon (7), Politics (8), Grok/xAI (6)
Posting Pattern	3 bursts with 3-4h dormant gaps

Output: output/musk-tweets-2026-02-09.json

Demo 2: WebAssembly 2026 Ecosystem Deep Dive

/deep-research "WebAssembly 2026 ecosystem"

The agent combined arXiv Advanced Search (academic papers) with Google (industry sources) to produce a comprehensive ecosystem report:

What the agent did:

Queried Actionbook API for arXiv Advanced Search selectors (40+ form fields)
Searched arXiv: title="WebAssembly", CS category, date range 2025-01 to 2026-02 → 20 papers found
Searched Google for industry sources (WasmCon 2025, Fermyon, The New Stack)
Deep-read 8+ sources via actionbook browser
Synthesized academic + industry perspectives into a single report

Report highlights:

Metric	Value
Spec Version	WebAssembly 3.0 (published Feb 6, 2026)
WASI Version	0.3.0 (WASIp3, imminent)
arXiv Papers Found	20 (CS category, 2025-2026)
Research Clusters	Serverless/Edge (8), Security (5), IoT (2), Tooling (3), Other (2)
Key Runtimes	Wasmtime, Wasmer, WasmEdge, WAMR

Key finding: Wasm is already everywhere — most users don't realize it. The academic consensus: Wasm excels at cold starts and workload density vs containers.

Output: output/webassembly-2026-ecosystem-v2.json

Why Actionbook?

Traditional AI tools (WebFetch, WebSearch) can only do simple keyword searches and read raw HTML. Actionbook is different — it indexes website UI structures and gives AI agents verified selectors to operate complex web forms.

Example: arXiv Advanced Search

Actionbook has indexed the entire arXiv Advanced Search form (40+ selectors). This means the AI agent can:

What the agent can do	How
Search by Title, Author, or Abstract separately	Select field via `#terms-0-field` dropdown
Filter to Computer Science papers only	Click `#classification-computer_science` checkbox
Restrict to papers from 2025-2026	Set date range via `#date-from_date` / `#date-to_date`
Add multiple search terms with boolean logic	Click "Add another term +" button

Example: X.com (Twitter) Search

The agent can log into X.com, use advanced search operators (from:user since:date), scroll through results, and extract engagement metrics — all via Actionbook browser automation.

None of this is possible with WebFetch or WebSearch — they can only send a single keyword query.

Quick Start (from Zero)

Prerequisites

Node.js 18+ (check: node --version)
A Chromium-based browser (Chrome, Brave, Edge, Arc)
An Anthropic API key

Step 1: Install Claude Code

npm install -g @anthropic-ai/claude-code

Verify:

claude --version

Step 2: Install Actionbook CLI

npm install -g @actionbookdev/cli

Setup:

actionbook setup

Actionbook is in Public Beta, you can use it without API_KEY.
If you want to increase the usage limit, join our waitlist.

Step 3: Install json-ui (Report Renderer)

npm install -g @actionbookdev/json-ui

This provides the json-ui CLI that converts JSON reports into self-contained HTML files.

Verify:

json-ui --version

Step 4: Add the Deep Research Skill

Option A: One-command install via npx skills (recommended)

npx skills add actionbook/actionbook --skill deep-research -g

This installs the skill globally — it works in any directory with Claude Code.

Option B: Install from local clone

If you've already cloned this repo:

npx skills add ./playground/deep-research -g

Option C: Manual copy

mkdir -p ~/.claude/skills/deep-research
cp playground/deep-research/skills/deep-research/SKILL.md ~/.claude/skills/deep-research/SKILL.md

Step 5: Run Your First Research

Start Claude Code:

claude

Then type:

/deep-research "WebAssembly 2026 ecosystem"

Or in natural language (supports English and Chinese):

Research the WebAssembly 2026 ecosystem and generate a report

That's it! The agent will search the web, read sources, generate a report, and open it in your browser.

Command Reference

/deep-research <topic> [options]

Parameter	Required	Default	Description
`topic`	Yes	—	Any topic, technology, URL, or `arxiv:XXXX.XXXXX`
`--lang`	No	`both`	`en`, `zh`, or `both`
`--output`	No	`./output/<slug>.json`	Custom output path

Topic Detection

Pattern	Type	Strategy
`arxiv:XXXX.XXXXX`	Paper	arXiv Advanced Search + ar5iv deep read
Academic keywords	Academic topic	arXiv Advanced Search + Google
URL	Specific page	Actionbook browser fetch and analyze
General text	Topic research	Google search + arXiv if relevant

More Examples

# Research a specific website's data (like X.com, Reddit, HackerNews)
/deep-research "Research how many tweets Elon Musk posted today on X.com"

# Deep dive a technology ecosystem
/deep-research "WebAssembly 2026 ecosystem"

# Analyze an arXiv paper
/deep-research "arxiv:2501.12599"

# Search by research topic (uses arXiv Advanced Search)
/deep-research "large language model agent papers 2025"

# Report in Chinese only
/deep-research "LLM inference optimization" --lang zh

# Custom output path
/deep-research "RISC-V ecosystem" --output ./reports/riscv.json

How It Works

┌──────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────┐
│  Claude   │────▶│  Actionbook  │────▶│  Web Pages   │────▶│ Extract  │
│  Code     │     │  Browser CLI │     │  (multiple)  │     │ Content  │
└──────────┘     └──────────────┘     └──────────────┘     └─────┬────┘
      │                                                           │
      │          ┌──────────────┐     ┌──────────────┐           │
      ├─────────▶│  Actionbook  │     │ arXiv Adv.   │           │
      │          │  search/get  │────▶│ Search Form  │──────────▶│
      │          │  (selectors) │     │ (40+ fields) │           │
      │          └──────────────┘     └──────────────┘           │
      │                                                           │
┌──────────┐     ┌──────────────┐     ┌──────────────┐           │
│  Open in │◀────│   json-ui    │◀────│  Write JSON  │◀──────────┘
│  Browser │     │   render     │     │  Report      │  Synthesize
└──────────┘     └──────────────┘     └──────────────┘

Workflow Steps

Step	Action	Description
1	Plan	Decide search strategy based on topic type
2	Query Actionbook API	Get verified selectors for known sites (arXiv, X.com, etc.) BEFORE browsing
3	arXiv Advanced Search	Multi-field academic search using Actionbook selectors (if topic is academic)
4	Google / Bing	Supplement with blogs, news, code, non-academic sources
5	Deep Read	Visit top sources, extract content via verified selectors
6	Synthesize	Organize findings into structured sections
7	Generate	Write a json-ui JSON report (bilingual EN/ZH)
8	Render	Produce self-contained HTML via `@actionbookdev/json-ui`
9	View	Open the report in your browser
10	Close	Close the Actionbook browser session

Features

Bilingual Reports (EN/ZH)

All reports are bilingual by default. The Chinese content follows quality guidelines to ensure natural, native Chinese writing — not machine translation.

Key principles:

Chinese text is written independently, not translated from English
Active voice preferred over passive voice
Technical terms with established Chinese names use Chinese; others keep English
Short sentences preferred (Chinese readers prefer concise phrasing)

Actionbook Browser Automation

The skill uses actionbook browser CLI commands exclusively (not WebFetch/WebSearch):

actionbook browser open <url>        # Navigate to page
actionbook browser text [selector]   # Extract text content
actionbook browser click <selector>  # Click element
actionbook browser snapshot          # Get accessibility tree
actionbook browser close             # Close browser

json-ui Report Components

Reports use 20+ @actionbookdev/json-ui components:

Component	Use For
`BrandHeader` / `BrandFooter`	Actionbook branding
`Section`	Major report sections with icons
`Prose`	Rich text content (Markdown)
`ContributionList`	Numbered key findings
`MetricsGrid`	Key stats dashboard
`Table`	Data tables
`LinkGroup`	Source links
`Callout`	Important notes / warnings
`CodeBlock`	Code snippets
`Formula`	LaTeX equations (for papers)

See skills/deep-research/SKILL.md for the full component catalog.

Troubleshooting

Issue	Solution
`actionbook: command not found`	`npm i -g @actionbookdev/cli`
`json-ui: command not found`	`npm i -g @actionbookdev/json-ui`
`claude: command not found`	`npm i -g @anthropic-ai/claude-code`
Browser won't open	`actionbook browser status` — ensure Chromium browser is installed
Empty report	Check internet connection, try a simpler topic
HTML render fails	The JSON report is saved at `./output/<slug>.json` — install json-ui and render manually: `json-ui render <file>.json -o <file>.html`
`[object Object]` in HTML	Update json-ui: `npm i -g @actionbookdev/json-ui@latest` (fixed in v0.1.0+)
Skill not found	Ensure SKILL.md is at `~/.claude/skills/deep-research/SKILL.md`
Chinese text reads like machine translation	The skill includes quality guidelines — re-run to regenerate

Dependencies

This skill requires two npm packages installed globally:

Package	npm Name	Purpose	Install
Actionbook CLI	`@actionbookdev/cli`	Browser automation, web search, selector queries	`npm i -g @actionbookdev/cli`
json-ui	`@actionbookdev/json-ui`	JSON report → self-contained HTML rendering	`npm i -g @actionbookdev/json-ui`

One-command install:

npm install -g @actionbookdev/cli @actionbookdev/json-ui

These dependencies are also declared in .claude-plugin/plugin.json under the dependencies and setup fields, enabling future auto-install support in Claude Code skill managers.

Project Structure

playground/deep-research/
├── .claude-plugin/
│   ├── plugin.json              # Plugin manifest
│   └── marketplace.json         # Marketplace metadata
├── .mcp.json                    # Actionbook MCP server config
├── skills/
│   └── deep-research/
│       └── SKILL.md             # Main skill definition (core logic)
├── commands/
│   └── analyze.md               # /deep-research command
├── agents/
│   └── researcher.md            # Research agent (sonnet, Bash+Read+Write)
├── examples/
│   └── sample-report.json       # Sample json-ui report
├── output/                      # Generated reports (gitignored)
│   ├── musk-tweets-2026-02-09.json
│   ├── webassembly-2026-ecosystem-v2.json
│   └── ...
├── .gitignore
└── README.md

License

Apache-2.0