Skip to content

Feature · Planning · Test design

It drafts.
You approve.

Describe a workflow in plain English and the agent writes a test plan you can read — steps and assertions in your language, not a script you have to take on faith. You approve what ships, and from then on it runs the same way every time.

plans you can readsuggestions ≠ requirementsno model at run time

The rules it cannot break

Not prompt wording.
Checks on the answer.

A prompt that asks a model to behave is a request. These are applied to whatever comes back, so a reply that ignores them does not get through.

A quote, or it is only a suggestion

An assertion may claim you asked for it only if it carries a quote that really appears in your description. The quote is checked against your own words; one that is not there is rejected and the assertion is recorded as a suggestion instead — kept and shown, but unable to fail anything until you adopt it.

Approval is not something the reply can contain

If the model returns a definition marked approved, the approval is dropped. Approval is a separate act by a person, and a plan that changes after approval has to be reviewed again rather than inheriting the old one.

Scope comes from you, never from the answer

Which workspace, which project, which environment and whether the run happens in the cloud or on your machine are taken from your configuration. If a reply tries to set them, the values in it are ignored.

Secrets never enter a definition

A literal password or token in a drafted step is replaced with a reference to a variable before the plan is stored, and a question that asks for the value of a secret is rewritten before you ever see it.

Unsupported ideas are dropped, not guessed at

A step or assertion the runner cannot perform is removed and named in the plan, rather than quietly translated into the nearest thing that happens to run.

Approved plans replay without the model

Once approved, the plan is stored and executed exactly as written. Nothing is re-generated at run time, so a test does not quietly change its mind between runs.

In the product

How it actually works.

1

Say what should happen

A sentence is enough. Ground it further with a requirement document or a browser recording and the plan is written against what is actually there rather than what the model expects.

2

Read the plan, not the code

Steps and assertions come back in plain language, each marked with where it came from: your words, the specification you attached, or the agent's own idea.

3

Adopt, refine or reject

Accept the suggestions worth keeping, refine the plan in a sentence and see exactly what changed, or throw it away. Nothing is saved until you approve it.

4

Run it, and keep running it

The approved plan executes in a real browser and goes on executing unchanged — in the app, from the CLI, or in CI — with no model in the loop at run time.

qamind — feature walkthroughVIDEO

See these steps on screen

coming soon

A guided video of this exact flow is being recorded — it will live right here.

Spec sheet

The fine grain.

InputPlain English · requirement documents · browser recordings
OutputA structured plan: steps, assertions and their origin
Origin checkQuote-verified against your own description
ApprovalHuman only — an approval in the model's reply is discarded
ScopeWorkspace, project, environment and run location come from your config
SecretsLiteral values replaced with references before storage
At run timeThe stored plan is replayed — no generation, no drift
WhereIn the app, in the CLI, or in CI

Read the plan before it becomes your test suite.