When a prompt scores low, the instinct is to rewrite the whole thing. That is the slow way, and it usually trades one weak prompt for a different weak prompt. The fast way is mechanical: find the single lowest load-bearing criterion, make one targeted change, re-score, and repeat. Below we take a real prompt from 2.1 to 4.0 in four passes, fixing exactly one criterion each time, and the last skill is knowing when to stop before you over-engineer it. This is the repeatable version of the loop we promised in the 3-minute rubric.
The mistake is chasing the overall number
A score is not a grade to maximize. It is a diagnostic that points at the weakest load-bearing part of the prompt. Treat it like a grade and two things go wrong.
First, you rewrite everything at once, change five variables, and cannot tell which change helped. You are back to debugging through a slot machine. Second, you chase 5 on every criterion, including the ones the task does not need, and you bloat the prompt with examples and reasoning scaffolds that make the answer worse, not better.
The discipline is the opposite of a full rewrite. Change one thing, the lowest load-bearing criterion, and measure. If it moved the number, keep it. If it did not, you learned something about the task. Either way you are never guessing.
Ordering matters because the criteria are not equal. Clarity, context, and structure carry double weight, so moving context from a 1 to a 4 buys more than pushing few-shot from a 1 to a 5, even though both are a four-point swing on paper. Fixing the lowest load-bearing criterion first is simply the highest-leverage edit available at each step. The full Prompt Score model has the exact weights, but the shortcut holds: start at the top of the load-bearing list, break ties left to right, and only move on once the number confirms the change.
The loop
Five steps, and only two of them involve writing.
Score the prompt on the six criteria (the rubric takes about three minutes).
Find the lowest load-bearing criterion. Clarity, context, and structure carry the most weight, so a 1 there costs more than a 1 on few-shot. Ties go to the one earlier in that list.
Make one targeted change that addresses only that criterion. Do not touch the rest.
Re-score. Confirm the number moved, and that nothing else regressed.
Save the version with its score, so the climb is visible later.
Your prompts can improve. Promptimizer rewrites and auto-tests them for you.
The prompt remediation loop: score, find the lowest load-bearing criterion, make one change, re-score, save
The loop. Each pass changes exactly one criterion, so you always know what moved the number. The save step turns a one-off fix into a visible history.
Four passes, from 2.1 to 4.0
Here is a prompt a small team actually shipped to a support inbox:
Reply to this customer asking for a refund.
Scored cold, it lands at 2.1: clarity 3, context 1, TCOF 3, role 2, few-shot 1, chain-of-thought 2 (top three weighted double). The lowest load-bearing criterion is context, at 1. So that is pass one.
Pass 1, fix context (1 to 4), score 2.8. The model has no idea what was bought, what the refund policy is, or why the customer is asking. We add it: "The customer bought the annual Pro plan 40 days ago and wants a refund because they forgot to cancel after the trial. Our policy allows a full refund within 30 days, a prorated refund after that." Nothing else changes yet. Re-scoring confirms that context alone lifted the whole prompt by seven tenths, which tracks: it was the single most underweight load-bearing criterion, so it had the most room to move the weighted total.
Pass 2, fix clarity (3 to 5), score 3.2. "Reply" is still vague: approve or decline, how long, what tone. We make it explicit: "Decline the full refund politely, offer the prorated amount, keep it to four sentences, warm and non-defensive."
Pass 3, fix TCOF structure (3 to 5), score 3.7. The prompt is now a run-on. We split it into Task, Context, Output, and Format so the model stops guessing at boundaries. Same information, four labeled blocks.
Pass 4, add a role (2 to 5), score 4.0. A generic assistant writes generic support copy. We name the role: "You are a senior customer support lead who protects the relationship without setting a policy precedent." That shifts the register in one line.
The score climbing from 2.1 to 4.0 across four passes, one criterion fixed each time
The techniques you're reading about work. Test your prompts now with Prompt Score and see your score in real time.
The climb. Four passes, one criterion each: context, then clarity, then structure, then role. Because only one thing changed per pass, every step of the number is attributable.
The finished prompt:
[Role] You are a senior customer support lead who protects the relationship without setting a policy precedent.
[Task] Reply to a customer requesting a refund. Decline the full refund politely and offer the prorated amount.
[Context] The customer bought the annual Pro plan 40 days ago and forgot to cancel after the trial. Policy: full refund within 30 days, prorated after that.
[Output] Four sentences, warm and non-defensive, ending with the prorated figure and a clear next step.
[Format] Plain email body, no subject line, no bullet points.
When a change does not move the number
Sometimes you fix the lowest criterion and the score barely budges. That is information, not a failure. Either the criterion was not the real bottleneck, in which case you re-score and move to the next lowest, or the task itself is underspecified in a way the rubric cannot see. A prompt asking the model to "predict which customers will churn next quarter" can score a clean 5 on structure and still fail, because the task needs data the prompt cannot carry. When a well-targeted fix stops moving the number, that is usually the signal the problem has left the prompt and moved into the task. Stop editing wording and go find the missing input instead. It is also the moment to ask whether this prompt needs an eval at all, or whether a structural score is still enough.
Where to stop
The finished prompt scores 4.0, not 5.0, and that is the right place to stop. Few-shot is still a 1 and chain-of-thought is still a 2, on purpose. A four-sentence refund reply does not need worked examples, and it does not need a reasoning scaffold. Pushing those two up would add length and lower the quality of the answer.
Before and after: the six criteria at pass 0 versus the finished prompt, with few-shot and chain-of-thought left low on purpose
Before and after. The three load-bearing criteria went to the top; role followed. Few-shot and chain-of-thought stayed low by design, because the task does not depend on them.
This is the part people miss. The goal is not a 5 on every line, it is a 5 on the criteria the task actually depends on. We stop when the load-bearing criteria are covered and the remaining lows are ones the task genuinely does not need. A refund reply that scores 4.0 with two deliberate lows is finished.
The stopping shape is task-specific, which is exactly why "get every criterion to 5" is the wrong target. A data-extraction prompt lives or dies on few-shot: two or three labeled input-output pairs move it further than any amount of role-play, so a good one might sit at 5 on few-shot and 2 on role. A multi-step analysis prompt needs the chain-of-thought a refund reply can skip. Read which criteria the task leans on, push those to the top, and leave the rest where they are.
Let the tool run the loop
The loop is mechanical, which is exactly what makes it worth automating. Scoring, finding the weakest criterion, and re-scoring after each change is repetitive work, and repetitive work is where consistency slips when you do it by hand at the end of a long day.
Keep My Prompts runs the loop for you: it scores the prompt, and Quick Optimize or Deep Optimize proposes the targeted rewrite, then re-scores so you can see the number move before you keep the change. Every version is saved with its score, so the climb from 2.1 to 4.0 is a visible history, not something you have to remember. On a team the saved climb also settles arguments: "the prompt went from 2.1 to 4.0" is a fact, not an opinion, and the next person can see exactly which criterion each pass fixed instead of relitigating the wording. If you would rather not start from a low score at all, the template library is full of prompts that already sit in the 4s, built to the same six criteria.
The discipline
Fixing a weak prompt is not about inspiration or a full rewrite. It is a loop: score, fix the lowest load-bearing criterion, re-score, save. Run it a few passes and stop when the criteria the task depends on are covered. The prompts that climb from a 2 to a 4 are not the ones someone rewrote in a flash of insight. They are the ones someone measured, changed by one criterion, and measured again.
