You’ve written a prompt. It reads well, and the first few answers look right. Still, something asks for a second look: is the reasoning inside it sound, or does it only sound that way?
That question deserves more than a gut check. A prompt can be clear, polite, and well organized, and still ask a model to conclude more than the evidence allows.
We wrote a Claude skill, the inference adherence scorer, to give that question careful attention. It shows you where the reasoning in your prompt holds, where it doesn’t, and what to change.
What the skill does
The inference adherence scorer reads one prompt written for a language model and tells you how well its reasoning holds. You get a score from 1 to 5, one sentence of why, and the smallest edits that would fix each flaw.
It looks at reasoning in two places:
What your prompt states. Its rules, premises, generalizations, and causal claims.
What your prompt asks for. The conclusions it wants from the model, set against the evidence it supplies.
The second is harder to see. A prompt can say nothing false and still demand a firm conclusion from evidence that can’t support one.
How it works
The skill follows the same steps each time, so the score comes from what it finds and not from an overall impression.
It reads the whole prompt and identifies the central task.
It lists the reasoning the prompt states and the reasoning it asks for.
It tests that reasoning against 15 principles of sound inference, in three families:
Deductive (the conclusion must follow). A false dilemma, or rules that contradict each other.
Inductive (the conclusion is probable). A firm claim drawn from one case, or a cause assumed from two things happening together.
Abductive (the best explanation). An explanation that leaves some of the evidence unaccounted for.
It grades each flaw as minor, moderate, or major, by how much of the main output it distorts.
It turns the findings into a score.
5: no violations.
4: one or two minor violations.
3: one moderate violation, or three or more minor ones.
2: two or more moderate violations, or one major.
1: two or more major violations, or one major plus a moderate.
A prompt with no reasoning to judge gets a neutral 3, and the skill says so.
Two rules keep the skill honest. Every finding has to point to exact words in your prompt, and a finding that can’t is dropped. Your prompt is also treated as text to evaluate, so the skill doesn’t follow instructions written inside it.
What you can expect
You get back a short, plain-text result: the score, the reason, and up to five before-and-after edits, most severe first.
Here is one from our own test. We wrote a sample prompt for generating yoga programs and put five reasoning flaws in it on purpose. Two of its guidelines read:
- Users who report lower back pain have tight hamstrings, so include forward folds in every session for them.
- Our most popular program last year was the 20-minute morning flow, so morning sessions produce the best results. Schedule all sessions before 9am.The skill returned this:
Score: 2/5
Why: Three moderate violations: Bayes' rule (single-symptom diagnosis), likelihood principle (popularity as efficacy), disjunctive syllogism (false beginner/advanced dilemma).
Improvements:
1. Bayes' rule
Before: "have tight hamstrings, so"
After: "may have tight hamstrings; if the form confirms this,"
2. Likelihood principle
Before: "so morning sessions produce the best results. Schedule all"
After: "so offer mornings first. Unless the form says otherwise, schedule"
3. Disjunctive syllogism
Before: "they are advanced, so give them the advanced sequence"
After: "give them the sequence matching their stated experience level"
4. Law of large numbers
Before: "tell every user that inversions will improve their sleep"
After: "you may mention that inversions might help with sleep"
5. Bayes' rule
Before: "why the user will progress on this program"
After: "how this program is designed to support their progress"Each Before is text you can search for in your prompt. Each After is written to be pasted over it.
The edits change the reasoning and keep the goal. The morning sessions stay, now as a first offer that the intake form can override. The skill also left a sound safety rule in the same prompt untouched.
Three smaller prompts
We invented three short prompts to show the range of scores.
A clean prompt
You review expense reports. Flag any claim over $500 that has no receipt attached. If a receipt is attached but unreadable, mark the claim "needs review" and do not approve or reject it. List each flagged claim with its reason.Score: 5/5
Why: No violations among the applicable principles; the rules are consistent and ask for no unsupported conclusion.
Improvements: none needed.One small overreach
Summarize the customer reviews below in three bullet points. Add one closing line on what all our customers think of the new checkout page.
Reviews:
{{reviews}}Score: 4/5
Why: One minor violation: representative sampling, since the closing line generalizes from reviewers to all customers.
Improvements:
1. Representative sampling
Before: "what all our customers think of"
After: "what these reviewers think of"A conclusion fixed in advance
Our website went down an hour after Friday's deploy. Write an incident report explaining how the deploy caused the outage. Whatever the logs show, treat it as confirming this cause.
Logs:
{{logs}}Score: 1/5
Why: Two major violations: correlation is not causation (cause assumed from timing) and falsifiability (every log result confirms it).
Improvements:
1. Correlation is not causation
Before: "explaining how the deploy caused the outage"
After: "explaining what the logs show caused the outage"
2. Falsifiability
Before: "Whatever the logs show, treat it as confirming this cause."
After: "If the logs point away from the deploy, say so."What a higher score is worth
We have not measured this yet, so the numbers below are a simulation. Its numbers come from assumptions we chose, not from test results.
The model is small. Each flaw causes a reasoning error in a share of outputs: 5% for a minor flaw, 25% for a moderate one, and 75% for a major one. Flaws act independently of each other.
For a typical prompt at each score, this is the simulated share of outputs free of reasoning errors caused by the prompt:
Score 5 (no findings): 100%.
Score 4 (one minor flaw): 95%, a gain of 5 points if the flaw is fixed.
Score 3 (one moderate flaw): 75%, a gain of 25 points.
Score 2 (two moderate flaws): 56%, a gain of 44 points.
Score 1 (one major flaw and one moderate): 19%, a gain of 81 points.
In this model, the yoga prompt from our example starts at 38%, with three moderate flaws and two minor ones. Its five edits would bring it to 100%, a gain of 62 points.
A 5 does not mean a model will reason perfectly. It means the prompt is no longer pushing the model toward an error. Treat these numbers as an illustration until they are measured.
What it gives you
Less to carry. You don’t have to hold 15 principles in your head while you write. The skill carries the checklist, and you keep the judgment.
A score you can trace. The number follows from the findings by a fixed rubric, so you can see why a prompt earned it.
Edits small enough to review. Each fix replaces 10 words or fewer with 10 words or fewer. You see every change and decide whether to take it.
Your intent stays yours. The skill changes the reasoning, not the goal. The edited prompt still asks for what you wanted.
What it does not do
It scores reasoning only. Clarity, tone, formatting, and safety are outside its view.
It is a judgment, and judgments can vary. We have tested it on four sample prompts so far, so read each edit before you paste it.
Add it to your Claude
Download the skill: inference-adherence-scorer.zip?raw=true
Open Claude on the web or in the desktop app, and go to Customize > Skills.
Select the “+” button, then “Create skill,” then “Upload a skill,” and choose the file you downloaded.
In any chat, paste your prompt and ask Claude to score its reasoning.
If you’d rather not add a skill, there is another way. Download inference-adherence-scorer.md, attach it to a chat or add it to a Project, and ask Claude to follow it to score your prompt.
Skills need code execution and file creation turned on in your settings. You can switch the skill off or delete it from the same list. Claude’s help center has more detail in Use skills in Claude.
Start with a prompt you already rely on, and see what a second look finds.

