BLUEPRINT BENCHOpen workbench

Skill packaging workbench

Package AI skills the way you'd release anything else.

A skill is a prompt plus guardrails, and it deserves the same rigour as any other piece of software: a spec, a review, a test suite and a version. Blueprint Bench takes a skill through all four before it reaches production.

Start a blueprint
  • ·No sign-up. Blueprints save to your browser.
  • ·Every score comes with reasoning, not just a number.
  • ·Status is earned from results, never toggled by hand.
PROMPTscope + instructionsGUARDRAILSwhat it must not doTEST CASESinput to expected behaviourREADINESS0-100 scorev0.2.0 MANIFEST
01

Define

Write the skill's prompt, guardrails and scope, the same brief you'd hand an engineer, plus a category and one-line description so it's easy to find again.

02

Audit

Gemini scores the blueprint 0–100 on clarity, safety, scope and completeness, then names the two or three risks a reviewer would actually raise, not a generic checklist.

03

Test

The skill runs against a real input under its own guardrails, then a second, independent model call grades the response against what you expected. Wording alone never passes.

04

Package

Version, status and readiness are set from the audit score and test pass rate, not a dropdown, then exported as a manifest a skills registry can ingest directly.

Drawing
Overview
Rev.
0.1.0
Status
ACTIVE
System
Blueprint Bench