Development · Tip
agent-experience Claude Code Skill
Audits documentation readability for AI agents by running parallel onboarding tests
A Claude Code skill that stress-tests developer documentation and SDKs by spawning parallel Claude subagents to simulate real-world onboarding. The agents attempt installation and quickstart workflows across multiple languages, identifying friction points where LLMs get stuck. It outputs an HTML audit report featuring an overall grade, narrative findings, prioritized fixes, and individual agent execution traces.
How it works
Flow shown in video: run /agent-experience in Claude Code → paste docs URL (example: https://docs.browserbase.com/welcome/introduction) → answer batched questions: parallel subagents (5/10), languages (TypeScript, Python), personas, exec mode, model (Sonnet/Opus/Haiku/mixed) → subagents attempt real onboarding (npm install, running quickstarts) → report: overall grade (Browserbase docs scored B 84/100), counts, narrative review, recommended fixes ranked by severity, what worked / what didn't, per-agent cards with tools, duration, status.
Findings example: hub pages lack runnable code; projectID should be explicit in quickstart; llms.txt underpromoted; playwright install chromium missing from Python quickstart.
Try at: https://www.skills.sh/browserbase/skills/agent-experience
What it's for
- Audit developer documentation to evaluate how easily AI coding agents can navigate it
- Identify missing dependencies, unclear quickstart steps, or broken code snippets across Python and TypeScript
- Benchmark docs before and after publishing agent-specific resources like llms.txt
Key points
- Invoked directly within Claude Code via the /agent-experience command
- Configurable test runs: set parallel subagent count (e.g., 5 or 10), target languages, and persona types
- Flexible execution models supporting Claude Sonnet, Opus, Haiku, or mixed setups
- Simulates real onboarding flows including package installations and quickstart script runs
- Generates detailed HTML audit reports with scores, severity-ranked fixes, and agent run traces
Pricing
Free to install and run; requires Claude API usage via Claude Code.
Details
- Type
- Tip
- Category
- Development
- Topic
- Developer Tooling
- Open source
- Yes
- Platforms
- Command line
- For
- Developer advocates, technical writers, and software engineers building agent-facing tools and docs
- Tags
- claude code, ai agents, documentation, dx, testing, benchmarking
Why it was saved
- @JaySahnan: Reusable workflow worth trying on any docs site — the video shows the full flow and the report format (narrative review, fixes, per-agent traces). Launch tweet claims ~30% faster agent onboarding internally.
Related
Command-line browser automation tool optimized for AI coding agents and testing.
Browser automation CLI and engine purpose-built for AI agents
Agent skills enabling AI coding agents to create programmatic Remotion videos
Open directory and CLI for discovering and installing AI agent skills
Curated library of specialized AI agent personas and workflows for Claude Code
Design vocabulary, rules, and skills for AI coding agents.
Updated