Sdd Flow — Testing skill for Claude Code

sdd-flow

[](https://github.com/zakyrion/sdd-flow/actions/workflows/ci.yml) [](LICENSE) [](package.json)

**Spec-Driven Development for AI coding agents.** `sdd-flow` installs a workflow into your project that makes agents like **Claude Code** and **Codex** treat every engineering request as a contract: normalize the request into an explicit spec, ask about everything that is unclear, wait for your explicit *go*, research before planning, plan before touching a single file — and record all of it in a persistent FLOW document that survives across sessions.

**You don't need to know Clojure.** The workflow writes specs in a tiny instruction notation that borrows Clojure's *syntax* because it is compact and unambiguous — nothing is ever executed, there is no language to install, and the agent both writes and reads it for you. You always answer in plain prose if you want to.

Why

Anyone who works with coding agents keeps hitting the same failure modes:

• The agent runs ahead. You describe a problem — it starts editing files.
• The agent invents scope. Unclear points get silently "resolved" instead of asked.
• The agent answers a question with a project. Ask how it should be done — get a migration roadmap, or the migration itself.
• The agent forgets. Close the session, and tomorrow's agent re-derives (or contradicts) yesterday's decisions.
• The agent walks in circles. On a long R&D task it re-proposes the approach that already failed last week — as a fresh idea.
• The agent answers from vibes. Ask about a genuine trade-off — terrain generation, netcode synchronization, service topology — and get a confident summary of its own training data, or the first article it found, handed over as the answer.
• **The agent writes the co