Agent Skill Tester for Markets

Anthropic's structured-skill spec: a markdown file with a name, description, input/output schema, and worked examples. Skills bundle a small repeatable agent capability (e.g., 'extract a 10-K risk factor') in a portable format. The tester loads your SKILL.md and runs it against sample input.

Calls go to Anthropic's API directly from your browser. The tool never sees or proxies the key. This keeps cost on your account, makes rate limits predictable, and avoids the privacy issue of routing finance prompts through a third-party proxy. Your key stays in browser memory only — it's not persisted to localStorage.

Three things: structured-output compliance (does the model return valid JSON matching the schema?), token cost (input + output tokens × current pricing), and end-to-end latency. Repeated runs show variance — useful for diagnosing flaky outputs.

The tester enforces strict-mode JSON schema validation. Common failures: missing required fields, nested objects with no properties defined, enum values outside the declared list. The validation error shows the exact field path. Real-world agents are more forgiving but also more inconsistent — strict-mode catches issues that bite later.

The current version is single-turn only — input → output. Multi-turn skill testing (where the model asks clarifying questions or takes multiple steps) is on the roadmap. For now, multi-turn skills need to be tested in a loop programmatically against the API directly.