“Run one more iteration” works best when the critique is specific. A score can supply that critique, as long as you know what the number measures and you decide in advance when to stop.
The pattern comes from email-coaching tools. Lavender, for example, documents a workflow in which a draft written with ChatGPT inside its extension is scored automatically, with recommended fixes and the potential score increase listed, and an option to apply the fixes to generate new copy (Lavender Help Center). Any email scorer can play that role, and so can a written rubric when you have no scorer at all.
What the number is, and what it is not
- A scorer’s score is that tool’s heuristic. It reflects the criteria the vendor chose. It is not an objective measure of whether this buyer will reply, so do not present it as one.
- An AI self-score needs a rubric. “Rate this email 1–10” with no defined criteria produces a number the model cannot explain consistently. Give it criteria and bands, and ask for the words behind each score.
- The reasons matter more than the total. A score is useful because it points at specific weaknesses. Read the flags, not just the number.
The loop
- Draft from evidence. List the facts the email may use and the facts that must not change: names, the verified signal, any buyer quote, the ask.
- Score it. Use your scorer, or the rubric below.
- Rewrite only what was flagged. Fix the lowest-scoring criteria and leave the protected facts alone.
- Re-score. Confirm the flagged criteria improved and nothing else got worse.
- Stop. Stop at your target or after two rewrite rounds, whichever comes first. Then a person reads it and decides whether to send.
A rubric when you have no scorer
Score each criterion 0, 1, or 2. This is a Promptifi working rubric for consistent critique; it does not predict reply rates.
- Accuracy (gate): 2 = every claim about the prospect traces to a fact you supplied; 0 = any claim does not. A 0 here fails the email whatever the total.
- Relevance: 2 = the opening connects one verified signal to a consequence for this role; 1 = the signal is mentioned but not connected; 0 = generic.
- Buyer focus: 2 = the buyer’s situation leads and the product is secondary; 1 = balanced; 0 = the email is about the seller.
- Ask: 2 = one specific, low-friction ask; 1 = an ask that is vague or oversized; 0 = no ask, or several.
- Length and scanability: 2 = readable at a glance on a phone; 1 = a paragraph too long; 0 = a wall of text.
Bands: 9–10 with Accuracy at 2 is ready for human review. 6–8 means rewrite the criteria that scored below 2. 5 or below means restart from the brief, because the problem is the evidence or the angle, not the wording. Set your target before the first score, not after.
Illustrative instruction — the scoring pass
“Score the email below against this rubric: [PASTE RUBRIC]. For each criterion, give the score and quote the words that earned it. Then rewrite only the criteria that scored below 2. Do not change these facts: [PROTECTED FACTS]. Do not add claims, urgency, or personalization I did not supply. Re-score the rewrite the same way. Stop when the total reaches [TARGET] with Accuracy at 2, or after two rounds, and tell me what is still weak. Email: [PASTE EMAIL]”
Stop rules
- Two rounds, then decide. Repeated polishing can sand off the specific language that made the draft credible.
- Never buy points with claims. If the score rose because the rewrite added urgency, a statistic, or personalization you cannot verify, revert it.
- You outrank the scorer. If a recommendation conflicts with what you know about this buyer, keep your version and note why.
- A good score is not approval. The person sending the email still reads it as the buyer would.
Put it to work
Improve a Sales Email and Get a Rewritten Version gives you a ten-point audit and a rewrite to score. Cold Email Review returns a 1–10 rating with its rewrite; paste the rubric above into it so the rating has defined criteria. Before anything goes out, Stress-test an AI-drafted outreach message before sending is the accuracy gate. For revising against explicit criteria more broadly, see Stop Asking AI to ‘Make It Better’ and Always run at least one iteration before you send.
