I moved coding-agent verification out of the prompt

Hermann Fischer — Fri, 21 Aug 2026 13:09:58 +0000

A coding agent can agree to run the tests and still fail to run them. It can also run the wrong command, overlook a dirty working tree, or declare success before the result is committed.

Adding another sentence to the prompt does not change who controls the completion decision. The same model doing the work still decides whether it followed the instruction.

I built Yoke to put that decision in a TypeScript runner.

Yoke executes work as a sequence of stories. For each story, the runner creates an isolated Git worktree, starts a fresh agent session, runs the configured verification command, requests a separate review, and checks that the resulting commit landed. A story is not marked as passed because the agent says it is done.

For user-interface work, the acceptance criteria can require Playwright screenshots or video. The evidence belongs to the story instead of living in a chat transcript.

Yoke supports Claude Code, Codex CLI, and Gemini CLI. A versioned canon generates the native instructions each agent expects, so a project can keep one methodology without pretending that all three tools use identical configuration formats.

Some of Yoke is Markdown. The skills, policies, and project context should be readable and editable. The enforcement layer is code: process supervision, worktree isolation, persisted run state, verification exit codes, review gates, and commit checks.

The closest existing project is WorkOS Case. Case uses a deterministic TypeScript pipeline to turn issues into reviewed pull requests with evidence. Yoke covers a somewhat different workflow: creating or retrofitting a project, converting a PRD into ordered stories, and running those stories across Claude, Codex, or Gemini. The projects share more philosophy than I realized when I first described Yoke, and I now call that out directly.

Yoke is MIT-licensed and does not require an account.

npm install -g @hecer/yoke
yoke new my-app

Repository: https://github.com/HECer/yoke

I would value examples of failures that prompts and skills do not handle reliably. I am also interested in the opposite feedback: which Yoke gate feels unnecessary in normal development?

Disclosure: I used AI assistance while drafting and editing this article. The product description and technical claims were checked against the repository by the maintainer.

DEV Community: Hermann Fischer

I moved coding-agent verification out of the prompt