Agent Skills, evaluated across every provider.
Author SKILL.md bundles per the agentskills.io spec. Six-dimension critique on the bundle. A multi-provider Behavioral Eval scores how reliably your skill activates and how consistently Claude, GPT, and Gemini read it. Publish the result as a public Skill Report. BYOK at every tier.
What ships with every workspace.
Free tier ships every AI instrument for Skills. Behavioral Eval per-run probe caps are uniform across tiers; the cost-preflight gate and rate limiter bound runaway-spend cases. There’s no markup on inference.
SKILL.md plus scripts/ and references/ in one bundle, edited in CodeMirror with inline frontmatter validation. Import from Anthropic, OpenAI Custom GPT export, or Gemini Gem export. The agentskills.io spec is the source of truth.
Discovery Fidelity, Instruction Quality, Example Coverage, Cross-Provider Portability, Token Efficiency, and Security & Safety Posture. The same shape as the prompts scorecard, retuned for capability bundles.
Run trigger and non-trigger probes across 2 to 5 BYOK models in parallel. The LLM judge scores activation accuracy, instruction adherence, and cross-provider consistency. Cost preflight gates the run; rate limiter caps runaway spend.
Frontmatter validation, file-layout discipline, prose hygiene, plus seven security rules that flag hardcoded secrets, untrusted fetch, runtime install commands, and other capability-specific risks. Runs on save and on demand.
Save any Behavioral Eval as a public Skill Report at /share/skill-report. Owner-revocable, robots-noindex by default, opt-in body publish. The same artifact backs a Shields.io badge for your GitHub README.
Snapshot-based version history with diff and restore, identical to prompts. Pull from your repo via the REST API or the TypeScript SDK. The agentskills.io 500-line body soft cap is enforced as a linter rule; bundles cap at 1 MiB per file and 4 MiB total.
The day-in-the-life of a Skill.
Six stations. Different from prompts: a Skill needs a lint pass and a Behavioral Eval, and it ships with a public Report and a badge. The workbench drives all of them.
- 01Author
Start a SKILL.md from scratch with the AI pair, or import an Anthropic SKILL.md, an OpenAI Custom GPT export, or a Gemini Gem export. Scripts and references attach as multi-file bundles.
- 02Lint
The 19-rule linter runs on save. Frontmatter shape, file layout, prose, plus the seven-rule security tier. Errors block the publish path; warnings stay visible until you address them.
- 03Critique
Score against the six skill dimensions. The same Workbench Model that drives Critique drives the rest of the AI pair.
- 04Improve
Apply targeted edits, accept the full Rewrite, Brainstorm variant approaches, or Compare two revisions of the same Skill. The six AI pair instruments work on Skills the same way they work on prompts · including Convert, which bridges a plain prompt into a complete Skill bundle in one shot.
- 05Behavioral Eval
Author trigger probes (cases that should activate the skill) and non-trigger probes (cases that should not). Run them across Claude, GPT, and Gemini in parallel. Read the activation and adherence scores per provider.
- 06Badge
Save the eval as a public Skill Report. Embed the Shields.io badge in your README. Pull the latest version via the REST API or the TypeScript SDK when you ship.
BYOK at every tier. Including free.
Behavioral Eval runs charge to your Anthropic, OpenAI, and Google accounts directly. Prompt Assay never proxies or marks up inference. Cost-preflight estimates the run before it starts; the rate limiter caps runaway-cost cases. Per-run probe caps (5 models, 10 trigger, 6 non-trigger) are uniform across every tier.
The platform demo budget for new signups does not cover Skills. You connect a BYOK key before running your first critique or eval. See tiers and cost for the cap details, or our trust page for data-handling.
Six questions we get every week.
- What is an Agent Skill, exactly?
- An Agent Skill is a multi-file capability bundle per the agentskills.io specification: a SKILL.md file with YAML frontmatter, optional scripts/ directory of helper code, and optional references/ directory of long-form context. Anthropic's Claude Code consumes the format directly; OpenAI Custom GPTs and Google Gemini Gems consume variants of the same shape. Where a prompt is one block of text aimed at one call, a Skill is a reusable capability that ships across providers.
- Why run Behavioral Eval across multiple providers?
- A Skill that activates correctly on Claude can fail to discover on GPT, or misread instructions on Gemini. Running trigger and non-trigger probes across Claude, GPT, and Gemini in parallel measures three things: activation accuracy (does the skill fire when it should and stay quiet when it should not), instruction adherence (does the model follow the SKILL.md once activated), and cross-provider consistency (do the three providers agree on the answer). Anthropic cannot ship a neutral cross-provider grader; Prompt Assay can.
- How is six-dimension Skills critique different from the prompt scorecard?
- The shape is the same (six dimensions, weighted scores, per-dimension feedback) but the dimensions are different. Prompts score on Clarity, Completeness, Structure, Technique Usage, Robustness, and Efficiency. Skills score on Discovery Fidelity (does the description make the model find the skill), Instruction Quality, Example Coverage, Cross-Provider Portability, Token Efficiency, and Security & Safety Posture. The dimensions were tuned against the agentskills.io spec and real-world skill failures.
- What does BYOK look like for Skills?
- Same as prompts. Your Anthropic, OpenAI, and Google API keys connect directly via BYOK. Behavioral Eval runs charge to your provider account, not to Prompt Assay. There's no markup on a single token. The platform demo budget for new accounts doesn't cover Skills; you connect a key before running an eval or a critique.
- Can I share a Behavioral Eval result publicly?
- Yes. Save any eval run as a Skill Report at /share/skill-report/[id]. Reports are owner-revocable, robots-noindex by default, and let you opt in to publishing the probe bodies. The same artifact backs a Shields.io README badge. See /docs/skills/badge for the embed pattern.
- Does MCP work with Skills today?
- MCP (Model Context Protocol) was donated to the Linux Foundation in late 2025 and is the cross-vendor protocol for AI tool integration. MCP support is on the Prompt Assay roadmap as the next workbench surface. Today, Skills ship as authored bundles per the agentskills.io spec; MCP integration is upcoming.
Author your first Skill in an afternoon.
Free to start. No credit card. Connect a BYOK key when you’re ready to run your first Behavioral Eval. The community anthropics/skills repository is a good reference set to read first.
Also shipping prompts to production? See the Prompts surface