Skillbench: the browser-based Claude prompt playground

A Claude prompt playground is a tool for running and iterating on Claude prompts without writing code. Skillbench is one such playground — free, browser-only, no account, and focused on the Skill artefact rather than a one-shot prompt. You write a SKILL.md in the form, run it against the live Anthropic API with your own key, and watch the output stream in real time next to a baseline run. Nothing leaves your browser except the API request itself, which goes straight to api.anthropic.com.

Skillbench vs. Anthropic Console

The Anthropic Console is the official prompt workbench. It's powerful, well-maintained, and the right tool for a lot of jobs — but it's optimized for ad-hoc prompt iteration inside an authenticated session, not for authoring and testing a Skill as a portable artefact. Where the two tools diverge:

The honest pitch: if you have an Anthropic account and you're iterating on a generic prompt, the Console is fine. If you want to write a Skill, share it as a URL with a teammate who doesn't have an Anthropic account, run it ten times to check variance, and commit the resulting .md file to a repo — that's the Skillbench shape.

Privacy and the BYOK model

Bring-your-own-key (BYOK) means your Anthropic API key stays in your browser. The default storage is sessionStorage — cleared when you close the tab. If you want the key to persist across tabs you opt in explicitly to localStorage, with a confirmation prompt. Skillbench has no server, no backend, no logging endpoint. Every API call goes from your browser directly to api.anthropic.com, which means Skillbench cannot see your prompts, your responses, your usage, or your key — even if we wanted to. The full source is on GitHub and the build is reproducible.

Three starter prompts you can fork

The gallery ships six Skills you can open with one click, fork, and run. Three to start with:

• code-reviewer.md

  Reviews diffs for bugs, security holes, and missing tests. Cites line numbers, ends with a ship / no-ship verdict.

  Open in Skillbench →
• email-triager.md

  Sorts an incoming email into respond / archive / delegate, attaches a one-paragraph draft for the respond bucket.

  Open in Skillbench →
• bug-repro-writer.md

  Turns a vague bug report into a minimal reproduction with numbered steps, expected output, and observed output.

  Open in Skillbench →

Each Skill loads as a complete SKILL.md — frontmatter and body — into the builder. Edit two lines, run again, watch the diff in the Compare mode. The other three Skills (research summarizer, Slack tone-shifter, meeting notes refiner) are linked from the gallery on the homepage.

When NOT to use Skillbench

Three cases where another tool is the right choice:

• Production multi-turn agents. If your prompt is part of a long-running agent loop that calls tools, branches on tool output, and accumulates state across turns, that workflow lives in Claude Code or an SDK, not in a single-shot browser playground.
• Retrieval-augmented generation. Skillbench does not load external corpora into context. If your prompt depends on a vector store of internal docs, test it where that store lives.
• Tool-call orchestration. The allowed-tools field in a SKILL.md declares which tools the Skill can use, but the tools themselves run in the agent runtime. Skillbench shows what Claude does with the Skill loaded as text; it does not execute file reads, web searches, or code blocks.

Inside those limits — single-shot Skill authoring, A/B comparison against a baseline, N-run variance measurement, share-by-URL — Skillbench is the fastest path from idea to versioned SKILL.md, with nothing to install and nothing to sign up for.