Clean spec in. Clean code out. FIX THE SPEC! AI agents build what the spec says, not what you meant. Speclint scores specs 0-100 across 5 dimensions (measurable outcome, testable criteria, constraints, vague verbs, verification steps). Below 70? It rewrites the failing parts automatically. CLI: npx speclint lint "your spec" GitHub Action: scores on issue open MCP Server: inline in Claude/Cursor REST API: for pipelines Open source scoring engine (MIT). Free tier, no signup.