Tool · Capstone

Claude Skills Builder

Three gates: Define "done" → Write the steps → Audit & update. Walk out with a one-page instruction sheet your AI follows every time.

What is a "skill"?

A skill is an SOP for the AI: a one-page document that says "every time I need to do X, here's exactly how: what 'done' looks like, the steps, the edge cases, the examples." Think of it like seasoning the fries the same way every time: write the recipe once and every batch comes out the same. You write it once, paste it at the top of any AI chat that's about that task, and the AI works exactly like the best version of you would. You teach it the way you'd train a new hire, with about 10 good examples and 2 bad ones, so it learns both what to do and what to avoid. No retyping. No drift.

Gate 1 · Define "done"

1. What this does for you

Before you can teach the AI to do a task, you have to know what 'done' looks like. Gate 1 forces that out of your head and onto the page in 4 questions.

2. What to type into your AI

Copy the box below. Replace the words in [BRACKETS] with your own info. Then paste it into Claude or ChatGPT.

I want to build a one-page instruction sheet I can reuse every time I need to do this task. Walk me through GATE 1: defining "done."

THE TASK I DO REPEATEDLY: [WHAT YOU FIND YOURSELF DOING OVER AND OVER, in your own field, e.g. "drafting the weekly update I send every client," "turning messy notes into a clean summary," "writing the same kind of proposal or report section," "answering a question customers or residents ask all the time"]

ASK ME THESE QUESTIONS, ONE AT A TIME, WAITING FOR MY ANSWER BEFORE THE NEXT ONE:

1. TRIGGER: When does this task start? What email, calendar event, file, or moment kicks it off?
2. DOCUMENT TYPE: What is the OUTPUT you produce? A doc? An email? A spreadsheet update? A categorization? Be specific.
3. OUTPUT FORMAT: How is it formatted? Sections? Bullet points? Length? Tone? Show me an example of what "done" looks like.
4. PUSHY CONFIDENCE: When you see this output, what makes you say "yes, that's it" vs. "no, redo"? List 3 must-haves and 2 must-NOT-haves.

After I answer all four, summarize back to me a one-paragraph "definition of done" I can reuse. Don't move to GATE 2 yet.

Try it now in Claude →

3. What to do with the answer

Pick ONE task you do over and over: the weekly client update, a recurring proposal or report section, turning notes into a clean summary, a reply to a question you get all the time. Whatever your field, pick the one that eats the most time. Don't try to skill-ify everything at once.
Run the Gate 1 prompt. Answer each of the 4 questions one at a time; don't batch them. The AI is doing useful work asking them in order.
When the AI gives you back a one-paragraph "definition of done," save it. That paragraph alone will save you arguments down the road about whether something is finished.

Gate 2 · Write the steps

1. What this does for you

Now turn how YOU do the task into numbered steps the AI (or a new hire) can follow. Plus the one weird edge case, plus a good example, plus a bad example.

2. What to type into your AI

Copy the box below. Replace the words in [BRACKETS] with your own info. Then paste it into Claude or ChatGPT.

Now GATE 2: methodology. We have a definition of done from Gate 1. Now we need the steps that get there reliably.

DEFINITION OF DONE: [PASTE WHAT YOU GOT FROM GATE 1]

ASK ME:
1. INPUTS: What information do I need to start? (Past examples, the org context document, the email that triggered it, etc.)
2. STEPS: Walk me through how I currently do this: every step, even the boring ones. I'll write them in order.
3. EDGE CASES: What's the one weird case where the normal steps don't work? What do I do differently then?
4. EXAMPLES: Give me one good example and one BAD example of the output. The bad one is more useful; it shows what to avoid.

After I answer, write me a numbered methodology under 150 lines that any new staff member could follow. Use plain language. No jargon. Don't move to GATE 3 yet.

Try it now in Claude →

3. What to do with the answer

Paste the definition of done from Gate 1 into the prompt.
Walk through the 4 questions. Be honest about your steps, even the boring "I copy-paste from last month's doc" steps. Those are the ones the AI will replicate.
Examples are how you teach the skill: aim to collect about 10 good examples and 2 bad ones over time, the way you'd train a new hire. The BAD ones matter more than they look: they tell the AI what to avoid. Start with one of each here and keep adding.
You'll get back a numbered methodology under 150 lines. Save it. This is the heart of your skill.

Gate 3 · Audit & update (come back in 2 weeks)

1. What this does for you

A skill that never gets audited drifts. Gate 3 is the 2-week check-in: did the methodology hold up? What broke? Update the skill so it stays accurate.

2. What to type into your AI

Copy the box below. Replace the words in [BRACKETS] with your own info. Then paste it into Claude or ChatGPT.

Final gate. GATE 3: audit and update.

WHAT WE HAVE SO FAR: [PASTE THE DEFINITION OF DONE AND METHODOLOGY]

I'm going to use this skill for two weeks. Then I want to come back here and have you ask me:

1. Did the methodology actually work? Where did it break?
2. What edge case did I hit that we didn't plan for?
3. What step turned out to be unnecessary? What step was missing?
4. Did the "definition of done" hold up, or did the real-world version drift?
5. Based on those answers, write me a v2 of the skill: same format, same length cap, with the changes baked in.

For now, give me a clean v1 of the skill in this format:

---
name: [skill-name]
description: [one line on when to use this and what it produces]
type: [reusable-prompt | sop | review-checklist]
---

# [Skill Title]

## When to use this
[trigger]

## What you'll need
[inputs]

## Steps
[numbered methodology]

## Edge cases
[the weird ones]

## Definition of done
[paste from gate 1]

## Example
[the good one]

## What NOT to do
[the bad example]

Try it now in Claude →

3. What to do with the answer

Use the v1 skill for at least 2 weeks. Actually use it. Notice when it works and when it doesn't.
Come back. Run Gate 3. Answer the 5 audit questions honestly, including "what step turned out to be unnecessary."
You'll get a v2. Replace your saved v1. Update the version number in the file name (skill-name-v2).
Set a calendar reminder for 90 days out to do this again. Skills decay if they're not maintained.

The capstone

By the end of this 3-gate flow you will have produced your first reusable skill: one document, under 150 lines, that turns a recurring task from a 30-minute scramble into a 3-minute paste-and-run. This is the goal of the entire training. Module 1-7 teach you how to talk to AI. Module 8 + this tool teach you how to teach the AI to be reliably useful, every time, without you re-explaining yourself.

Download the template

Starter skill template

Download a blank SKILL.md you can fill in as you walk through the three gates. The Markdown version is what you paste at the top of any AI chat. The PDF version is what you print for board or staff onboarding.

Section Recap

Core Principle

If you're going to do something more than 3 times, build a skill, not a one-off prompt. A skill is an SOP for the AI, a 1-page recipe that seasons the fries the same way every time without re-explaining.

Best-Fit Example

A busy professional writes the same kind of document over and over: a weekly client update, a recurring proposal section, a standard reply. She turns it into a skill. Gate 1 nails "done" (three specific must-haves, two must-nots). Gate 2 numbers the steps and adds a bad example, the time it drifted into vague filler; over a few weeks she grows it to about 10 good and 2 bad examples. Gate 3, two weeks in, bakes in the one rule she kept forgetting. The task that used to cost her half an afternoon now takes a few minutes, and reads sharper.

A teacher writes the same kind of report-card comment every term. She turns it into a skill. Gate 1 nails "done" (three sentences per student, evidence cited, no canned praise). Gate 2 numbers the steps and adds a bad example, the time a comment drifted into "a pleasure to have in class"; over the term she grows it to about 10 good and 2 bad examples. Gate 3, two weeks in, bakes in the rule to pull the strength straight from the rubric row. An evening of 25 comments now takes 30 minutes and reads sharper.

A city analyst writes the same council briefing memo before every meeting. She turns it into a skill. Gate 1 nails "done" (background, options, fiscal impact with a funding source, staff recommendation, neutral tone). Gate 2 numbers the steps and adds a bad example, the memo that got sent back for missing a funding line. Gate 3, two weeks in, bakes in the rule to define every acronym on first use. The memo that ate an afternoon now takes minutes and stops coming back for edits.

A superintendent writes the same weekly owner update from field notes. He turns it into a skill. Gate 1 nails "done" (work completed, work planned, any schedule slip in days behind baseline, open RFIs). Gate 2 numbers the steps and adds a bad example, the update that buried a delay. Gate 3, two weeks in, bakes in the rule to tie every slip to a cause. Fridays got shorter and the Owner stopped calling to ask what "behind" means.

An owner answers the same kind of customer review every week. She turns it into a skill. Gate 1 nails "done" (warm, signed by first name, one real next step on a complaint, no coupons unless she says so). Gate 2 numbers the steps and adds a bad example, the reply that sounded copy-pasted. Gate 3, two weeks in, bakes in the rule to flag reviews she should answer personally. A week of replies now takes ten minutes instead of an hour.

A development lead writes the same monthly donor update. She turns it into a skill. Gate 1 nails "done" (one number, one real story, one specific ask, no hype). Gate 2 numbers the steps and adds a bad example, the update that spun a month the program fell short. Gate 3, two weeks in, bakes in the rule to name any goal they missed. The update stays honest, and a task that took an afternoon now takes minutes.

A bookkeeper sends the same month-end summary to a dozen clients. She turns it into a skill. Gate 1 nails "done" (revenue vs last month, cash position, one thing to watch, labeled general guidance and not filed advice). Gate 2 numbers the steps and adds a bad example, the note a client read as if cash were profit. Gate 3, two weeks in, bakes in the rule to state cash vs accrual. The close note that took an afternoon now takes a few minutes per client.

An artist writes the same first reply to commission inquiries. She turns it into a skill. Gate 1 nails "done" (warm, no jargon, two questions about size and timeline, a starting price range). Gate 2 numbers the steps and adds a bad example, the reply that underpriced the work. Gate 3, two weeks in, bakes in the rule never to quote a full price before she knows the scope. Her reply time dropped to a couple of minutes, and she stopped leaving money on the table.

Key Takeaways

Pick ONE recurring task: pick the one that wastes the most of your time, not the easiest one.
Don't skip Gate 1. Vague definitions of 'done' are how skills die.
Teach the skill with examples: about 10 good and 2 bad. The bad ones teach the AI what to avoid.
Cap the skill at 150 lines. Anything longer is two skills pretending to be one.
Set a 90-day calendar reminder before you close the tab; Gate 3 only happens if you schedule it.

← All Tools