Claude Code · Educational Series

Teach Claude the why, not just the what

Anthropic's alignment research found that explaining the reasons behind good behavior generalizes better than showing examples of it. The same idea makes your prompts and skills far more reliable.

Reference Guidewith a downloadable resource for your Claude Code sandbox

What's inside

  1. What the research actually found
  2. Why this matters for your prompts
  3. Add a 'why' line to your skills
  4. What you'll notice
Section 01

What the research actually found

In a 2026 post, Anthropic described training Claude to behave well in tricky situations. Training only on demonstrations of the right action helped a little. But when they rewrote the same data to include the reasoning — why one choice was better than another — performance jumped, and it generalized to situations the training never covered.

Strikingly, a small dataset that taught principles matched the results of a much larger one built on examples — roughly a 28x efficiency gain. The takeaway: the reasons matter more than the actions.

The one-line summaryTeaching the principles behind good behavior beats training on demonstrations of it alone.
Section 02

Why this matters for your prompts

You're not training models, but the same dynamic shows up every day. A prompt that's only a list of steps works until reality differs from your script — then Claude follows the steps literally and produces something wrong.

When you include the goal — why you're doing this and what a good outcome looks like — Claude has something to reason from when it hits a case your steps didn't anticipate. The intent transfers; a bare recipe doesn't.

Plain versionSteps tell Claude what to do. The why tells it what to do when your steps don't quite fit.
Section 03

Add a 'why' line to your skills

Put the purpose at the top of any skill or prompt, before the steps. One or two sentences: what this is for and what good looks like. Then list the instructions underneath as usual.

This costs you a single line and changes how Claude handles everything that follows. It's the cheapest reliability upgrade you can make.

## Why this exists
Give the reader a clear weekly picture so they can
plan next week. Clarity matters more than completeness.

## Steps
1. ...
Try itTake a skill you already use, add a two-sentence 'why' at the top, and rerun it on a weird input. Watch the judgment improve.
Section 04

What you'll notice

With the intent stated, Claude makes fewer brittle, overly literal mistakes and shows better judgment on the edge cases — the inputs that don't match your examples. You'll find yourself writing fewer exceptions and special cases, because the principle already covers them.

Do both: state the why and give the steps. Anthropic found the combination is the strongest of all.

Don't overdo itOne clear sentence of intent beats three paragraphs of philosophy. Say the point, then get to the steps.
Free Download

The principle-first skill template

A SKILL.md template that puts the 'why' first, so Claude reasons from intent — based on Anthropic's research finding.

Why-first structure
Goal statement
Steps section
Edge-case note
Definition of done
Output example
Before/after demo
When to use
Research note
Checklist
✓ Downloaded. Drop it in your skills/ folder and try it in your Claude Code sandbox.