Skip to content

How I Turned 13 Years of Teaching High School AP Stats into a Claude Code Skill

How I turned 13 years of AP® Statistics teaching notes into a Claude Code skill that walks through basic statistical inference problems using the 4C Method.

AP® is a registered trademark of the College Board, which was not involved in the production of, and does not endorse, this skill or this post.

The Problem

Thirteen years of teaching AP Statistics leaves you with a library. Condition-checking templates. Calculator quick references. Inference frameworks and scoring rubric notes. The kind of material you accumulate one semester at a time, refining what works, discarding what does not.

The problem is where all that knowledge lives: in PDFs, in handwritten notes, in my head. Ask a generic AI to solve a stats inference problem and you get a textbook answer. It does not know the 4C Method. It does not know you should always open your conclusion with a p-value interpretation. It has no idea which TI-84 command to use or what the scoring rubric actually rewards.

The Solution: A Claude Code Skill

Claude Code has a feature called skills: markdown files that teach the AI domain-specific workflows. I built one called /stathelp that encodes how I actually teach basic statistical inference. Not how a textbook explains it. How I explain it, after thirteen years of watching students get it right and get it wrong.

The process was straightforward. I fed Claude my original teaching notes and let it extract the patterns. The source material:

  • My 4C Method templates for confidence intervals and significance tests (Choose, Check, Calculate, Conclude), based on the curriculum from StatsMedic that I used in my classroom for years
  • A TI-84 calculator reference I wrote covering every function students need on the exam
  • My notes on p-value interpretation as a teaching priority, emphasizing that students should always interpret the p-value in their conclusion, even though the exam does not require it for full credit
  • An inference procedure selection guide with the flowchart I use in class: Is it one of the weird ones? Means or proportions? One or two samples?

Claude extracted the teaching logic and organized it into a structured skill with reference files: a curriculum map covering all nine units and a deep inference guide with conditions tables, common student errors, conclusion templates, and the right calculator command for every procedure.

What It Actually Does

When I type /stathelp and give it a problem, the skill kicks in and Claude walks through it the way I would teach it:

  1. CHOOSE: Identifies the procedure, defines the parameter in context, writes hypotheses with the larger group first (so the test statistic comes out positive), and states the significance level.
  2. CHECK: Verifies all conditions with actual numbers shown: Random, 10%, and Large Counts (using the pooled proportion for significance tests, the sample proportion for confidence intervals).
  3. CALCULATE: Shows the general formula, plugs in numbers, reports the test statistic and p-value, and includes the TI-84 command.
  4. CONCLUDE: Starts with the p-value interpretation (assuming the null is true, the probability of the observed result or more extreme, purely by chance), then compares to alpha, states the decision, and concludes about the alternative in context.

After the solution, it flags common student errors specific to that problem type, things like using the sample proportion instead of the null proportion for the Large Counts condition, or writing hypotheses about statistics instead of parameters.

Grading Itself Against the Rubric

I gave it a real free response question: the 2019 FRQ #4 about kochia plants and glyphosate resistance. Then I fed it the official scoring guidelines and asked it to grade its own work.

It walked through each rubric section. Hypotheses and procedure identification. Conditions and calculations. Conclusion. It correctly identified which components it satisfied, earning a 4 out of 4. Complete response, all three sections essentially correct.

One thing it got wrong on the first pass: it labeled the parameters backwards, putting the 2014 proportion first and producing a negative test statistic. Mathematically equivalent. Still accepted by the rubric. But not how I would teach it. I told the skill to always put the expected larger group first. Now it does.

Why This Matters

This is not about replacing teaching. It is about encoding teaching expertise into a tool that can be reused. The 4C Method, the p-value interpretation emphasis, the specific errors I have watched students make for thirteen years. None of that lives inside a generic AI model. It lives in the notes I wrote and the patterns I refined semester after semester.

A Claude Code skill is just a markdown file. Fill it with domain expertise from someone who has actually done the work, and it becomes something more useful than any textbook answer. It becomes a teaching assistant that thinks the way you think.

The whole process took one conversation. Raw PDFs in, working skill out, good enough to pass an exam rubric. The skill keeps getting better as I feed it more resources. Thirteen years of notes, finally in a format that scales.

The skill is open source: github.com/schl0ss/claude-skill-stathelp