3.4 The 80/20 Iteration Loop

Learning Objective

Learn the practical iteration loop for refining skills from "good enough" to "production quality" - and understand why shipping at 80% and iterating beats chasing perfection before launch.

The 80/20 Reality

You can get to 80% accuracy surprisingly fast. Sometimes in 30 minutes. The remaining 20% takes 80% of the effort. This is not a guess - it is a pattern you will see over and over.

Most people never ship because they are stuck polishing that last 20% before they even start using the skill. Do not be that person. Ship at 80%. Use it in your real work. Improve it as you go. This is how every real system gets built.

The Loop

The iteration loop has five steps, and you will repeat it until your skill is consistently producing great results.

Build - write the skill. You already did this in Lesson 3.3.

Test - run it at least 5 times with different inputs. Not once, not twice. Five times minimum.

Review - look at each output critically. What is good? What is off? What is missing? What showed up that you did not want?

Refine - update the skill based on what you found. Add examples, tighten instructions, fix gaps.

Test Again - run it 5 more times with different inputs to verify your changes actually worked and did not break something else.

Repeat until you are hitting 90%+ consistency across varied inputs.

Why 5 Tests, Not 1 or 2

One test tells you almost nothing. You might have gotten lucky, or you might have gotten unlucky. Either way, a single result is not a pattern.

Five tests with different inputs reveal real patterns. You will see whether the output is consistent across runs, whether it handles different types of inputs, and whether it follows all the rules you set.

Use varied inputs when testing. Give it an easy case, a typical case, a tricky case, and an edge case. If you only test with perfect inputs, you will not know how it handles real-world messiness.

What to Look For When Testing

Consistency - does it produce similar quality every time, or is the output random? If you get a great result followed by a terrible one, the skill needs tighter instructions.

Edge cases - what happens with unusual inputs? Empty data? Weird formatting? Missing fields? These are the situations that will trip you up in production.

Output quality - does the output actually match the format and example you specified? Or is Claude drifting away from your template?

Missing steps - did it skip anything? Did it add things you did not ask for? Both are signs that your process steps need to be clearer.

Tone and style - does it sound like your business, or does it sound like generic AI output? If the tone is off, add style guidelines to your rules section.

How to Refine

Add examples. This is the single most effective fix for almost any quality issue. Show Claude exactly what "good" looks like, and the output quality jumps immediately.

Tighten instructions. Replace vague language with specific direction. "Write a good intro" becomes "Write a 2-sentence intro that mentions the client's industry and the specific problem discussed on the discovery call."

Add quality checks. Include a verification step: "Before finishing, verify that the output includes X, Y, and Z." This catches common omissions.

Handle edge cases. Add conditional logic: "If the input does not include a budget, ask for it before proceeding." This prevents the skill from guessing when it should be asking.

Add constraints. Specific limits like "Output must be under 300 words" or "Always use short paragraphs, never walls of text" keep the output consistent and usable.

When to Stop Iterating

Stop when the output is consistently good across 5 or more varied tests. Stop when you would be comfortable showing the result to a client or team member without editing it first.

If the remaining issues are minor style preferences rather than structural problems, you are done for now. You can always come back and refine later as you use the skill in real work. Do not get stuck in an endless refinement loop.

Action Items

• Run each of your skills at least 5 times with different inputs
• Document what worked and what did not for each test
• Refine your skills based on what you found - add examples, tighten instructions, add quality checks
• Run 5 more tests after refining to verify improvements
• Note your biggest insight from testing and share it in the community