Answer questions across 6 dimensions of your system. The AI synthesizes your answers into a structured AGENTS.md spec — fully private, runs entirely in your browser.
First load downloads the model (~1–2 GB). Cached in your browser after that.
Copy this into Claude Code, Cursor, or any AI agent to generate your project.
Grade a spec the way a coding agent will read it: 40 practices, 10 dimensions, a score out of 30, from “The EPIC Framework for Spec-Driven Development”. In the paper's 114 repos, top-scoring specs went with far fewer bug-fix commits (11.8% vs 20.4%). Nothing leaves your browser.
No spec built yet. Answer the questions in the Spec Builder and click Generate Spec, then score it here — or use one of the other tabs.
Your ZeeSpec spec is ready ().
Finds spec-kit (specs/*/spec.md), Kiro (.kiro/specs), ZeeSpec (.spec/*.md), AGENTS.md, PRDs, plans and task lists.
Same fictional product, written two ways.
Rules find evidence fast; a local LLM (WebGPU) re-reads each dimension against the paper’s 0–3 rubric and suggests one fix. Final score = rounded mean of rules and judge. Uses the builder's model if one is already loaded.
A spec is only as good as the checks built from it. Can a model turn one requirement into a harness that rejects a buggy implementation without rejecting a correct one that just behaves differently where the spec is silent? 16 tasks across the 8 verification-logic types of SpecGuardBench. Every harness runs in a sandboxed worker in your browser.
Each task has one atomic spec, a reference implementation and a test that already passes. Hidden from the author: a correct implementation, a buggy one that breaks the spec but still passes the existing test, and a valid alternative that satisfies the spec but differs wherever the spec is silent (error messages, ordering, extra fields…).
A harness is faithful only if it rejects the buggy implementation and accepts both correct ones. Detection = buggy one rejected. False alarm = a correct one rejected.
The mirror trap shows the buggy implementation as the reference. In the paper, agents then wrote checks that blessed the bug: