Writing an evaluation grid and having Copilot apply it

1Why a criterion must be observed, not judged

Anthropic's prompt engineering guide highlights an essential rule: define a success criterion before drafting your prompt. The evaluation grid provides this objective benchmark. It lists five observable criteria, each scored from 0 to 2 with a one-sentence justification. It applies to any version of the deliverable.

A solid criterion is observed immediately without subjective judgment. Here are five practical examples:

• The target audience and requested decision appear in the opening line.

• Every stated figure refers to a named and verifiable source.

• No sentence exceeds twenty-five words.

• The final recommendation specifies who does what and by when.

• Disagreements are isolated from approved decisions.

Why not simply ask Copilot to review its own work? Research by Huang shows that language models do not self-correct reliably without external feedback. Without an independent anchor, their performance often drops after an unguided self-correction attempt.

Reliable verification requires an element outside the initial generation: an attached source document or a predetermined rubric. Two strict rules follow: the rubric is written beforehand, and the evaluation runs in a separate conversation containing only the text to evaluate.

The workflow follows four stages. First, you draft the rubric. Second, you generate version 1 of the deliverable. Third, in a fresh conversation, you assign Copilot a reviewer role along with the rubric and the raw text.

You require a score per criterion, a concise justification, and three priority corrections formatted as instructions. You apply valid suggestions, reject ungrounded edits, and rerun the rubric on the revised draft. Microsoft's Think Deeper mode reinforces the analytical depth of this audit.

Tap to enlarge

2A board note that goes from 5 to 9 out of 10

An operations manager asks Copilot to write a one-page briefing proposing an organisational change to senior leadership.

Weak prompt, in the same conversation:

Is it good?

Copilot responds that the memo is "clear and well-structured" and suggests "strengthening the introduction". Nothing is measurable, nothing is actionable, and the memo retains two unverified figures.

Strong prompt: Turn 1 (independent audit in a new conversation):

You are a quality reviewer. Evaluate the memo below using this rubric. For each criterion, assign a score from 0 to 2 and a one-sentence justification:
1) Every figure refers to a named source;
2) The target audience and requested decision appear in the first sentence;
3) The final recommendation states who does what and when;
4) No sentence exceeds 25 words;
5) Foreseeable objections are listed with a response for each.
End with the three priority corrections, phrased as rewriting instructions. Do not rewrite the memo.
Memo: [pasted text]

Copilot applies the grid strictly. The verdict returns an overall score of 5 out of 10: two figures without sources (0/2), decision missing from opening paragraph (1/2), four sentences exceeding length limits (0/2).

Strong prompt: Turn 2 (human judgment and second audit):

I added the source for productivity gains and shortened the flagged sentences. However, I kept the union risk analysis that you suggested deleting, as management requires full transparency.
Here is version 2. Re-evaluate it using the same rubric and compare scores.
Memo V2: [pasted text V2]

The second audit reaches 9 out of 10. The sources are validated, the opening decision is sharp, and the manager's chosen tradeoff is acknowledged as well-grounded.

What changes: observable criteria produce a comparable verdict between iterations. Actionable instructions enable immediate revision. The manager's explicit review confirms that human oversight governs the final deliverable.

3Write five criteria checkable in one minute

Choose a routine deliverable from your daily work: client email, executive summary, meeting minutes, or product sheet. Here is the initial prompt to improve:

"Read my note and tell me what you think."

First, write your grid of five observable criteria, scored from 0 to 2. Produce version 1 with Copilot. In a new conversation, set up the reviewer role and submit the raw text with your rubric.

Request a score, a justification per criterion, and three actionable corrections without full rewriting. The composer below structures this prompt. Enable the booster to flag critical points for human verification. Apply relevant edits, discard questionable suggestions, and compare before-and-after scores.

Self-assessment grid: (a) each of the five criteria can be checked in under a minute; (b) the evaluation ran in a separate conversation; (c) scores before and after are recorded, and at least one edit was refused with a clear rationale.

Open the prompt composer