Getting started

How Diffie works

Every test Diffie runs, regardless of where it was triggered, flows through the same five stages. Knowing which stage a test is in makes it much faster to diagnose a failure or tune a test that keeps picking the wrong element.

  1. 01

    Plan

    The Diffie plugin in Claude Code analyzes your source (full repo or a diff) and proposes concrete test specs: steps, assertions, and what secrets they need. You approve each spec before anything is created in Diffie.

    Where you see it

    In Claude Code or in the PR-testing workflow log.

  2. 02

    Explore

    Diffie opens your app in a real Chromium browser and scans the DOM for selector-stable hooks: data-testid, aria-label, accessible role, and name. Tests target those hooks rather than fragile CSS paths.

    Where you see it

    Happens server-side on Diffie. Visible in the run recording and the selector log.

  3. 03

    Generate

    It emits Playwright code that follows the spec and targets the selectors it found. Diffie stores the generated code in your workspace so you can inspect, edit, or re-run it.

    Where you see it

    The test detail page in the Diffie dashboard.

  4. 04

    Run

    The test executes end-to-end with video recording, per-step timing, and network-request logs. Secrets referenced by name are injected at run time and never leave your workspace.

    Where you see it

    The run detail page, which includes a shareable public URL.

  5. 05

    Report

    Pass/fail is surfaced wherever you triggered the run: Claude Code, a PR comment, or the dashboard. Failing tests can be reprocessed with a short fix prompt, which loops back to stage 01 with extra context.

    Where you see it

    Claude Code chat, the PR comment thread, or the dashboard.

Reprocessing vs. regenerating

A reprocess re-runs stages 01 to 04 with an extra instruction you give in the fix prompt. The spec is updated, Diffie re-explores only what changed, and the generated code gets patched rather than rewritten from scratch. That's how auto-fix keeps tests stable across UI changes: it nudges the existing test instead of throwing it out.

Use a brand-new test (start over at stage 01 with a fresh prompt) when the flow you want to cover is genuinely different from what the existing test captures.

Next: author tests in Claude Code

The Diffie Plugin guides cover interactive test authoring end-to-end: installation, planning, selector hints, secrets, auto-fix, and common gotchas.

Diffie Plugin guide →