browse-and-evaluate（浏览并评估）

Browse And Evaluate

Goal

Find the right skill for a task without flooding the context window or installing blindly.

Guardrails

• Always use --fields on list/search/info to keep output small. Default: --fields name,tier,workArea,description.
• Always use --dry-run before installing anything.
• Never install more than 3 skills at once without explicit user confirmation.
• Prefer --format json in non-interactive pipelines. The CLI defaults to JSON when stdout is not a TTY.
• Use --limit when browsing large catalogs. Start with --limit 10.

Workflow

1. Search or browse the catalog.

npx ai-agent-skills search <query> --fields name,tier,workArea,description --limit 10

1. Get details on a candidate.

npx ai-agent-skills info <skill-name> --fields name,description,tags,collections,installCommands

1. Preview the skill content.

npx ai-agent-skills preview <skill-name>

1. Dry-run the install.

npx ai-agent-skills install <skill-name> --dry-run

1. Install only after reviewing the dry-run output.

npx ai-agent-skills install <skill-name>

Gotchas

• The preview command sanitizes skill content to strip prompt injection patterns. If content looks truncated, check if suspicious patterns were removed.
• Collection installs pull multiple skills. Always --list or --dry-run a collection before installing.
• Upstream (non-vendored) skills require a network fetch at install time. Use --dry-run to verify the source is reachable.