Score Any Prompt in 3 Minutes: The 6-Criteria Rubric We Run Before Hitting Send

You can tell a good prompt from a bad one before you ever run it. We do it in about three minutes, with a six-line rubric we keep in our heads. Most people do the opposite: they fire the prompt, read the answer, and if the answer is bad they blame the prompt. That is the wrong feedback loop. The output is noisy. The model got lucky or unlucky, the temperature moved, the same prompt gives you a different answer an hour later, so a single response tells you almost nothing about the prompt itself. Scoring the prompt's structure, before you run it, is faster and more repeatable. Below is the exact rubric we use, one criterion at a time, with the fast tell that separates a 1 from a 5, plus a worked example you can copy.

Why we score before we run

The habit came out of frustration. We would tweak a prompt, run it, get a mediocre answer, tweak it again, run it again, and after twenty minutes we still could not say what had actually changed. We were debugging through a slot machine.

There are three reasons judging a prompt by its output does not work. It confuses correlation with causation: a decent answer can come from a weak prompt when the model has rich training data for the topic, and a strong prompt can produce a weak answer on an inherently ambiguous task. It is not repeatable: unless you pin the temperature to zero, the same prompt returns different text every time, so evaluating a prompt from one response is like judging a die from one roll. And it does not scale: once you are reusing more than a handful of prompts, you cannot run and read every one to decide which needs work.

The fix is the same move code review made decades ago. A reviewer does not say "the code is good because the program ran." They read the structure: is it clear, is it complete, does it handle the edges. A prompt has structural properties too, and they are visible in the text without running anything. That is what the rubric measures.

The 3-minute rubric

Six criteria, each scored 1 to 5. For each one we ask a single question and look for one tell. You do not need to overthink the number. The point is the diagnosis, not two decimal places.

The six criteria, with the fast tell that separates a 1 from a 5. Score the top three first: they carry the most weight.

1. Clarity and specificity. Would a stranger know exactly what to produce? A 1 is a topic ("write about marketing"). A 5 has an action, a subject, constraints, and a success criterion ("write a 600-word marketing plan for an organic-cosmetics store, budget 3k a month, focused on Instagram and email"). If you can imagine two people reading it and building two different things, it is not a 5 yet.

2. Context. Could the model do this well without knowing your situation? A 1 gives it nothing. A 5 supplies the audience, the domain, the goal, and the relevant constraints, the information the model cannot infer on its own. Watch out for filler context: a generic "you are helpful" line adds words, not signal.

3. TCOF structure. Are Task, Context, Output, and Format all present? A 1 is a single undifferentiated blob. A 5 articulates all four. This is the criterion people skip most, and it is the one that most reduces the randomness of the answer. If you have never separated the four, the TCOF framework is worth ten minutes.

4. Role. Is there a specific role that shifts the model's register? A 1 assigns none. A 3 is generic ("you are an expert"). A 5 names real expertise and a perspective ("a senior data analyst who works on B2C seasonality"). The more specific the role, the more it narrows the model's response distribution toward the knowledge you actually want.

5. Few-shot examples. Does the prompt show, not just tell? A 1 has no examples. A 5 has two or three representative input-to-output pairs. This one is conditional: it matters enormously for non-standard formats and domain-specific classification, and barely at all for open creative writing. A low score here is not automatically a problem.

6. Chain-of-thought. For a reasoning task, is there a required thinking structure? A 1 has none. A 3 is the bare "think step by step." A 5 lays out the steps ("first identify the anomalies, then compare against the benchmark, then conclude"). Also conditional: essential for multi-step reasoning, pointless for a summary or a tweet.

Here is the only weighting you need to remember without a calculator: the first three criteria are load-bearing, the last three are situational. When we total the six in our heads, we let clarity, context, and structure count roughly double. A prompt that is a 5 on chain-of-thought but a 1 on clarity will reason beautifully about the wrong task. The reverse, clear but no explicit reasoning, still works for most jobs. If you want the exact weights and the evidence behind each criterion, the full Prompt Score model lays out the formula and the research.

A worked example, timed

Take a prompt we actually caught ourselves about to send:

Summarize this article for our blog readers.

Three minutes on the clock.

• Clarity: 2. "Summarize" how long? A paragraph, five bullets, a tweet? No angle, no length, no success criterion.
• Context: 1. Which article? Who are "our blog readers", and what do they already know? The model is guessing at everything that matters.
• TCOF: 2. There is a task. Context, output, and format are all missing.
• Role: 1. None assigned.
• Few-shot: 1. No example of a summary we liked before.
• Chain-of-thought: 1, but this one is a free pass. A summary does not need a reasoning scaffold, so we do not hold it against the prompt.

The sample prompt scored across all six criteria. With the top three weighted double, it lands near the bottom of the range. The V-Meter reads red, and the two lowest load-bearing criteria, context and clarity, are exactly where to start.

Weighting the top three double, this lands at roughly a 1.6 out of 5. That number on its own is not the useful part. The useful part is that the two criteria dragging it down are the load-bearing ones, context and clarity, so we know precisely what to fix first and we have not run the prompt even once. Rewriting from the lowest criterion up is its own discipline, and we cover the repeatable version of that loop in From 2 to 4: fixing a low-scoring prompt one criterion at a time.

Reading the total, and when a low score is fine

The score is a diagnostic, not a grade. A 1.6 is not a moral failing. A throwaway prompt for a quick brainstorm can score a 2.5 and be completely fine, because you are in the loop, reading and correcting every answer. The same 2.5 on a prompt you are about to reuse a hundred times, or hand to a teammate, or drop into an automation that runs unattended, is a liability. Score relative to the stakes and the reuse, not against some absolute bar.

The other trap is inflation. The two 1s in the example, few-shot and chain-of-thought, are not defects for a summary task. Padding the prompt with examples and a reasoning scaffold it does not need would raise the number and hurt the result, because context quality beats context quantity and models genuinely do get lost in the middle of long prompts. Do not chase five out of five on every line. Chase five on the criteria the task actually depends on.

When the manual rubric stops being enough

We still run the rubric by hand, and you should learn it by hand too, because scoring a dozen prompts manually is how the criteria become instinct. But the manual version breaks on three things.

Consistency: we score the same prompt a little higher on a fresh morning than on a tired Friday. Scale: once the library is dozens or hundreds of prompts, we are not scoring each one by hand. Memory: we do not remember what last month's version scored, so we cannot see whether a prompt is actually improving or just changing.

That is the point where we let the tool carry it. Keep My Prompts scores every prompt you save on these same six criteria, applies the same weights every time regardless of what day it is, shows the result on a color V-Meter, and saves the score with each version, so you can watch a prompt climb from a 2 to a 4 across its history. Same rubric, minus the fatigue and the forgetting. It is free to start, and if you are still deciding where prompts should live, the prompt management tools comparison covers the options.

Start with the manual rubric to build the instinct. Automate it for consistency across moods, scale across a library, and a saved score snapshot on every version.

Start manual, automate the reflex

The rubric is the point, not the tool. Run it by hand a dozen times and the six questions turn into a reflex you apply in the pause before you hit send. Then automate that reflex, so it never depends on whether it is Tuesday morning or Friday at six.

Either way, the shift that matters is treating a prompt as something you can measure before you run it, not something you can only judge after. The best prompts are not the ones that got lucky on the first answer. They are the ones that were built to score well before anyone pressed enter.