Specs your AI agent
can actually follow.
SpecStudio Skills is a free Claude Code plugin that turns vibe-coded prompts into structured, scored specs your agent reads before it writes a single line of code. Built on the openSpecScore protocol.
$ /plugin marketplace add specscore/ai-marketplace $ /plugin install specstudio@specscore
$ specscore agent setup <agent-name>Writes the right instruction file for your agent —claude, copilot, cursor, codex,antigravity.google, pi.dev, or opencode. Run inside a SpecScore-managed project. Requires theSpecScore CLI.
$ git clone https://github.com/specscore/specstudio-skills $ cp -r specstudio-skills/commands /path/to/your/agent/commands/ $ cp -r specstudio-skills/skills /path/to/your/agent/skills/
Clone the plugin repo and copy the commands/ andskills/ directories into wherever your agent loads command and skill files from.
From vibe to scored spec, in 60 seconds.
> /specstudio:ideate A password reset flow for our SaaS app. Refining idea... Producing spec/ideas/password-reset.md > /specstudio:specify spec/ideas/password-reset.md Writing Feature spec... spec/features/password-reset/README.md $ specscore lint spec/features/password-reset/ PASS Feature: password-reset PASS has title, status, and description PASS requirements have acceptance criteria PASS all AC IDs are unique PASS no orphan cross-references Score: 92 / 100
No edit. No cuts. Blank Claude Code → install plugin → score a real spec → agent ships the right thing.
Why your agent keeps building the wrong thing.
Vibe-coded prompts work until they don't. The agent built half the feature. It misread the requirement. It invented an API call. It picked the wrong library. You re-prompt. You re-prompt again. By the third round you're typing the spec out longhand into the chat box anyway.
The problem isn't the agent. It's that "build me a feature" is not a spec — it's a wish. Agents are stochastic. Wishes are stochastic-friendly. Specs are not.
SpecStudio Skills gives you a workflow for writing specs your agent can actually parse: structured features, requirements, acceptance criteria, scored for completeness before the agent ever sees them. Same speed as vibe-coding, dramatically lower variance in the output.
Four commands. One pipeline.
Install once, then run them in order — from raw idea to executable plan.
Refine a raw idea through structured divergent and convergent thinking. Produces a lintable pre-spec one-pager.
Turn an approved idea into a SpecScore Feature with requirements, acceptance criteria, and Given/When/Then blocks.
Decompose a feature into an ordered list of tasks. Each task maps to one or more acceptance criteria.
Execute the next task in the plan. One subagent per task, AC-traceable commits, staged diff for your approval.
What's in the box.
Does:
- Adds slash commands for the full SDD workflow:
/ideate,/specify,/plan,/implement. - Scores spec completeness and ambiguity before the agent sees the spec.
- Writes specs in the open SpecScore format — no proprietary extension.
Doesn't:
- Run the agent for you. The plugin is a workflow layer; the agent is still Claude.
- Phone home. Specs and scores stay local.
- Lock you in. If you uninstall, your specs are still valid Markdown and still scoreable by the specscore CLI.
Looking for the neutral CLI surface?ai-plugin-specscoreis the thin wrapper. SpecStudio Skills is the opinionated workflow on top.
Built on the open SpecScore protocol.
SpecStudio Skills writes specs inSpecScore, an open standard. The format is human-readable Markdown. The schema is published. The CLI is open source. No vendor lock-in. If you ever leave SpecStudio, your specs come with you.