Skip to content

Framework

What would make this good enough?

Seven questions for a decision. They help you show your reasons, not create common sense or decide for you.

  1. I

    Frame the problem

    Before looking at outputs, name the purpose, the people affected, and what 'good' means. Check the question too. A model’s answer and your first instinct both need checking.

  2. II

    Set criteria

    Write the standards first: accuracy, fit for purpose, fairness, tone and craft. This helps you judge the output against the task, not the other way round.

  3. III

    Check evidence

    Separate checked facts from claims, including your own assumptions. Fluent text and polished images do not prove correctness.

  4. IV

    Weigh context and stakes

    The same flaw is minor in a mood board and serious in a dosage note. Ask who could be harmed and whether the harm can be undone.

  5. V

    See tradeoffs

    Speed against care, consistency against detail, coverage against privacy. Say what each choice gives up.

  6. VI

    Test reliability

    Run it again. Vary the input. Compare with known facts where they exist. One good output says little about the next one.

  7. VII

    Decide, with reasons

    Accept, revise, hold or reject. Record why, so others can review, challenge and learn from the decision.

Passable is not enough

Passable output reads well and stays on topic. Better output fits the purpose, passes expert review and respects the people affected. Knowing the difference takes expertise and effort.

Review fatigue

More output means more to check. If tired or repeating a correction, pause and return to the standard. Smaller batches and a fresh qualified review may help.

Aesthetic judgment

In creative work, taste means seeing whether the work holds together and fits the intent. It means noticing when the work becomes generic. You can recognise the right direction without mastering the software. Explain the standard. Let a qualified person challenge it.

Authority and responsibility

Whoever approves an output answers for it. The reviewer need not be the engineer running the agents. People affected should be able to question the standard.

Why no score

A number can summarize a judgment. It cannot make one. A score can look exact while hiding who decided.

Check what works and why