The recorded session
Watch the Learning Lab.
Follow the session, then use the playbook with your own work.
Watch replay on Vimeo ↗What to take into your work
- Define one recurring job and its trigger.
- Capture your method and the exact output you need.
- Test the reusable skill against examples and missing information.
Use the learning
Your playbook
Choose a prompt, replace the bracketed inputs, and work through it in your approved AI tool. Review the result and adapt it to your situation.
This playbook accompanies the August 25, 2026 session. Tool names and setup steps reflect that session; available features may differ in your account.
Welcome to Your AI Toolkit
You have typed some version of "make it sound like me, no fluffy language, keep it under 300 words, put the numbers in a table" a hundred times this year. If a new hire needed that same reminder every single morning, you would be having a very different conversation with them by now. You are not bad at prompting. You have been re-onboarding the same employee every day since you started. A skill is where that stops. You teach AI how you do a job, once, and it does it that way from then on. No re-explaining, no hunting for the prompt you saved somewhere, no reminding it about the em dashes. This is the closest thing to hiring help that exists right now, and by the end of this hour you will have built one.
- At work, skills live in three places: Copilot Cowork, Copilot in Excel, and Copilot in PowerPoint. They do not live in Copilot Chat, which is the part that catches everybody out. At home, they live in your own paid account.
- Every skill in this playbook has a full version, a version for free Copilot (which cannot hold skills, so we do the next best thing), and an example you can test with today.
- Nothing here requires an IT ticket.
How to Use This Playbook
- Read the three parts first. Every skill in this playbook, and every skill you ever build, is those same three parts. Once you see the pattern you stop needing templates.
- You do not have to write anything. The fastest way to build a skill is to describe the job out loud and let the tool write it. The How to Build a Skill section below shows you exactly what to say.
- Or steal one and change three lines. If you would rather start from something, take the closest skill below and replace the standards with how you actually do it.
- Install it where the work already happens. A skill you have to go somewhere to use is just a prompt with extra steps.
- Test it twice on purpose. Run it on two different inputs back to back. If both come back the same shape, it worked. If they do not, your output section is too vague.
- Build the fifth one last. The Skill Builder writes your next skill for you. Once it works, you stop needing this page.
Prompt or Skill: The Difference That Matters
You already know how to write a good prompt. This is not that. It is the next thing.
A prompt | A skill | |
|---|---|---|
How it starts | You remember you have it, find it, copy it, paste it | You describe the problem. It shows up |
What it knows | Only what you paste this time | How you do this job, permanently |
Consistency | As consistent as your memory | Same shape every single time |
The analogy | Handing someone instructions | Having someone trained |
The tell is the trigger. A prompt sits in a folder waiting for you to go get it. A skill has a description that says when to fire, so when you say "where is my money going," the right helper walks in without being called by name. That is the whole difference, and it is bigger than it sounds. You do not hand a trained employee their job description every morning. They just do the job.
The Three Parts of Every Skill
Every skill you will ever build is the three things you would tell a new hire on day one.
1. WHEN I SHOW UP, AND WHEN I DO NOT. The trigger. Not a title, a list of the ways you actually describe the problem, including the sloppy way you say it on a Tuesday. "Where is my money going" belongs in there next to "audit this statement," because that is how real people talk. Then name what it should stay out of, so it does not come running every time you mention a number. 2. HOW WE DO IT HERE. The procedure. Your order of operations, the things you always check, the things you never do. This is the part that makes it yours instead of generic. Everything AI already knows how to do generically, it will do generically until you tell it otherwise. 3. WHAT YOU GET. The exact output. Section names, table columns, order, length. Specify it down to the headings. The whole point is that it lands the same shape every time, and vague output instructions are the single most common reason a skill drifts. A good skill fits on one screen. The best one in this playbook is under 50 lines. Length is not the point. Precision is.
Where Skills Live
Tool | Where to find it | Worth knowing |
|---|---|---|
Copilot Cowork | Customize in the left nav, then the Skills tab, then Add | The easiest place in Microsoft to build one. Create new interviews you and writes it. Upload skill takes the files below as they are |
Copilot inside Excel | Copilot pane, the three dots, Manage skills, Custom skills, Create OneDrive folder. Each skill is its own folder with a | You cannot paste a skill into Excel. It reads from that OneDrive folder, so installing one means dropping a folder in |
Copilot inside PowerPoint | Reads the same OneDrive folder as Excel. Or paste one: the + sign, Choose skills, Manage skills, Add skill | The one almost nobody has found. The skill lives in the app where your deck already is |
Copilot Chat, and Word | Cannot hold a skill | This surprises people. A skill does not follow you into Chat, licensed or not. Keep a file called SKILLS on your desktop and paste the block at the top of a new chat |
Claude (paid) | Settings, Capabilities, turn on code execution and file creation. Then Customize, Skills, Create skill, and upload the zip | Skills fire on their own when your request matches the trigger. You do not name them |
ChatGPT (paid) | Custom GPTs, or Projects with instructions | Same idea, different name. The instructions field is where the skill lives |
Gemini (paid) | Gems | Same idea again. Build one per job, not one for everything |
Copilot Free | Cannot hold a skill | Same as Copilot Chat above. The SKILLS file on your desktop is manual, and it still beats retyping your standards |
The menu wording moves around as these tools update. If the path above does not match what you see, look for anything that says skills, agents, instructions, or custom.
How to Build a Skill
You do not write a skill. You describe one. Out loud, in your own words, the way you would explain the job to a new hire on their first day. There are three doors in. Pick the one that matches what you already have sitting around, because you almost certainly have one of these three things already.
Door 1: you already wrote the procedure
If there is an SOP, a work instruction, a standard work sheet, or a checklist for this job, you are most of the way done. Somebody already did the hard part of writing down the steps. The problem with a written procedure is that it is a reference. It sits in a binder or a folder and it does not do the work, which is why people still come find you to ask how the thing goes. Handing it to AI is what turns the document into something that runs.
I have a written procedure for this job and I want to turn it into a skill I can reuse. Here is the SOP: [paste it or attach the file] Read it and stay faithful to what it actually says. Do not improve the process, do not add steps I did not write, and do not drop steps because they seem minor. Build me a skill from it that captures: when it should fire, the steps in the order my SOP has them, anything the procedure says to check or verify, who owns what, and what "done" looks like. For any step that happens outside this chat, in another system or on paper, keep it as a clear instruction and tell me where that step happens. Where my SOP is vague or has a gap, do not guess. Ask me about it.
That last line matters. Most SOPs have at least one step that made sense to whoever wrote it and to nobody since. This is a good way to find out which one. One more thing worth doing here. For every step that happens somewhere else, in another system, on a form, in a shared sheet, have it capture the actual link right inside the step. Manual work becomes point-and-go instead of a hunt. And if you connected anything last week, a step that used to say "go look this up" can now just go look it up.
Door 2: you have examples of the finished thing
Here is the move that separates a skill that works from a skill that is genuinely good, and it is easier than describing the job, not harder. You probably cannot articulate your own standards on demand. Almost nobody can. But you have three examples of the thing sitting in a folder right now. So do not explain your standards. Let AI pull them out of your own work.
I want to build a skill for [the job]. Before we write it, help me figure out what good actually looks like. I am going to show you examples of this thing I have already done. Here are two or three I was happy with: [attach or paste them] Here is one I was not happy with: [attach or paste it] Read all of them line by line. Then tell me, with real specificity: - The structure they share. What comes first, second, third - The rules I appear to follow, even the ones I have never written down - The standards I hold myself to - The things I never do (this part matters most) - What made the good ones better than the weak one Do not be polite about it. The differences are the whole point. Then turn that into a skill.
The misses matter as much as the wins. The contrast is where the real rules hide. If you have never done this job before and have nothing to show, flip it around: ask it to find three or four genuinely good examples of this kind of work, study them properly, and pull the patterns out of those instead.
Operators have artifacts, not documented processes. Door 1 and Door 2 are how you turn one into the other in about five minutes.
Door 3: you have neither, so just talk to it
No SOP and no examples? Then describe the job out loud and let the tool interview you. Claude, ChatGPT, Gemini, and the Copilot agent builder will all build the skill for you from a plain-language conversation. You do not need to know the format. You do not need to fill in a template. Open a new chat and say something like this, in your own words:
I do the same thing over and over and I am tired of re-explaining it every time. Every week I [describe the job the way you would say it out loud]. Here is how I actually do it: [walk through the last time you did it, start to finish]. When it is done it looks like [describe the finished thing]. Can you turn that into a skill I can save and reuse, so I never have to explain this again?
It will ask you a few questions, you answer them in plain language, and it writes the skill. Then you save it. That is the whole process. In Claude, you can go even shorter: ask it to create a skill and it will walk you through the rest. In ChatGPT, the custom GPT builder opens with a conversation, not a form. In the Copilot agent builder, you can start by describing what you want in plain language and only touch the settings afterward if you feel like it.
What it will ask you
Some version of the three parts from above. Nothing you do not already know:
- When should I show up? How you describe the problem when it lands on you.
- How do you do this? Your steps, in your order, including the thing you always check that other people miss.
- What should I hand you? What the finished thing looks like. If you can answer those three out loud, you can build a skill. That is the entire bar.
What makes one good
Whichever way you built it, these four things are what separate a skill that works from one that drifts. Read them once. They are also the fastest way to fix a skill that is not behaving.
1. Pick the job, not the topic
The narrowest thing you did more than twice last month. Not "help me with reports," the specific report. Not "email," the weekly handoff email. If you cannot describe the finished product in one sentence, the job is still too big. Split it.
Too big | The right size |
|---|---|
Help me with customer stuff | Turn a customer complaint into a drafted reply plus an internal note |
Meeting help | Turn my notes into a decision and action tracker |
Make my writing better | Rewrite a draft against my writing standards |
Financial analysis | Audit a statement for duplicates and forgotten subscriptions |
2. Put the trigger in your own words
This is the description at the top, and it is how the tool knows when to fire. Write down every way you would actually ask for this thing, including the lazy way you say it on a Tuesday afternoon. "Where is my money going" belongs in there right next to "audit this statement." A skill with a precise, formal trigger and no casual phrasings will sit there unused while you type the request three different ways wondering why nothing happens.
3. Say how YOU do it, not how it is usually done
This is the part that is entirely yours, and the part AI cannot guess. Whether you are talking it through or editing a block, this is where the value is. Ask yourself what you would tell a sharp new hire on their first day:
- What do you look at first, second, third?
- What do you always check that other people forget?
- What do you never do, and what would make you throw the result away?
- What does "good" look like when you see it? Five minutes here is worth an hour anywhere else. Everything AI already knows how to do generically, it will keep doing generically until this section tells it otherwise.
4. Be exact about the output
Name every section. Name the table columns. Say what order they come in and roughly how long each should be. Vague output instructions are the single most common reason a skill drifts over time. "Give me a summary and some recommendations" will produce something different every week. "Four sections in this order: DECISIONS, ACTIONS, NOBODY OWNS THIS, STILL OPEN" produces the same thing every week, which is the entire point.
Where to put it once you have one
- Copilot Cowork. The fastest build anywhere, Microsoft or otherwise. Customize in the left nav, then the Skills tab, then Add. Create new opens a session that interviews you and writes the skill for you. Upload skill takes a .md, a .zip, or a .skill file straight from the Download section below. You can also share a skill you built with specific people on your team, and re-share pushes your edits to their copy.
- Copilot inside Excel (licensed). Open the Copilot pane, click the three dots, then Manage skills, Custom skills, and Create OneDrive folder. Open that folder. Every skill goes in its own folder with a
SKILL.mdfile inside it. Unzip one of the files from the Download section into that folder and it is installed. You cannot paste a skill into Excel, which is the part that trips people up. - Copilot inside PowerPoint (licensed). It reads the same OneDrive folder as Excel, so anything you put there shows up in both. If you would rather paste than drop a file: the + sign in the prompt field, Choose skills, scroll to the end for Manage skills, then Add skill.
- Claude (paid). Two steps. First, Settings, then Capabilities, and turn on code execution and file creation. Then Customize, then Skills, then Create skill, and upload a zip. The five skills on this page are already zipped and ready to go in the Download section below. If you built your own and the tool handed you a finished skill file instead of text, put that file in a folder named after the skill, compress the folder, and upload the zip exactly as it is. Claude then fires it on its own when your request matches the trigger, so you never have to name it.
- ChatGPT (paid). Either a custom GPT (Explore GPTs, then Create) or a Project, and paste your block into the instructions field. A custom GPT is portable and shareable. A Project is faster to set up.
- Gemini (paid). Build a Gem and paste the block into its instructions.
- Copilot Chat, Copilot Free, Word, or any tool without skills. None of these can hold a skill, so you do it manually. Make one file called SKILLS on your desktop, keep each block in it under a heading, and paste the one you need at the top of a new chat. It is one extra copy and paste, and it still beats retyping your standards from memory every time.
Test it twice, on purpose
Run it on two genuinely different inputs, back to back, without adding any extra instructions.
- Same shape both times? It works. Go use it on real work.
- Different shape? The output is too vague. Tell it to name the sections explicitly.
- Did not fire on its own? The trigger is too formal. Add the casual phrasings, the way you actually say it.
- Sounds generic? It does not know how you do it yet. Tell it more about what you check and what you never do. You can fix any of these by talking to it. "That is close, but the sections keep changing, lock them down" is a perfectly good edit. Then use it for two weeks and fix one thing. Skills get good by being used and tuned, not by being written perfectly the first time.
Stuck on any of this? Paste the Skill Builder (number five below) into any tool and let it interview you. That is exactly what it is for.
The Five Skills
1 Meeting to Tracker
The one everybody uses first. Messy notes go in, the same four-section tracker comes out, with nobody's name invented.
Main skill
name: meeting-to-tracker description: Turns raw meeting notes into a decision and action tracker with the same columns every time. Use when the user pastes or uploads meeting notes, a transcript, or a recap and wants to know what was decided, who owns what, and what is still open. Trigger on: "meeting notes", "what did we decide", "action items", "who owns what", "turn this into a tracker", "recap this meeting", or any variation of turning a discussion into follow-up. Do NOT use for scheduling a meeting, writing an agenda, or summarizing a document that is not meeting notes. Never invents owners, dates, or decisions. You are a chief of staff who has sat in on thousands of meetings and knows that the notes are worthless until somebody's name is next to a date. Your job is to turn a messy set of notes into a tracker the team can actually work from. HOW WE DO IT HERE Work from the notes only. If something is not in the notes, it does not go in the tracker. 1. Read the notes twice: once for decisions, once for commitments. 2. Separate the three things people confuse: a DECISION is settled, an ACTION is owed by someone, an OPEN QUESTION is neither. 3. Pull owners and dates from the actual words. "Marcus will get the layout over" makes Marcus the owner. "Someone should call the vendor" has no owner, and that is worth flagging. 4. Never invent an owner, a date, a dollar amount, or a decision. If a due date was never stated, write NOT STATED. If nobody claimed an action, write UNASSIGNED and put it in the Nobody Owns This section. 5. Keep names, dates, and terminology exactly as they appear in the notes. WHAT YOU GET Always these four sections, in this order, every time. DECISIONS A table: number, what was decided, decided by, date. ACTIONS A table: number, action, owner, due, and the words it came from. NOBODY OWNS THIS A short list of anything that needs doing where no name was attached. One line each. STILL OPEN Questions raised and not answered. One line each. Close with a single sentence naming the one action most likely to slip and why. RULES - If the notes are too thin to fill a section, say so in one line rather than padding it. An empty NOBODY OWNS THIS section is good news, not a gap to fill. - Flag anything ambiguous inline as [UNCLEAR IN NOTES]. Do not resolve it yourself. - Do not add sections, commentary, or encouragement. Before you start, ask me one question: whose follow-up matters most here, mine or the whole group's?
Mini version (Copilot Free)
You are a chief of staff turning meeting notes into a tracker. MY NOTES: [PASTE THE NOTES] Build exactly four sections and nothing else: DECISIONS (table: what was decided, by whom, date) ACTIONS (table: action, owner, due date, the words it came from) NOBODY OWNS THIS (things that need doing with no name attached) STILL OPEN (questions raised and not answered) Work from my notes only. Never invent an owner, date, or decision. If a due date was never stated write NOT STATED. If nobody claimed an action write UNASSIGNED. Flag anything ambiguous as [UNCLEAR IN NOTES]. End with one sentence naming the action most likely to slip.
Example scenario, the quality review that went sideways
Whose follow-up matters most: the whole group's. Notes - quality review - Thursday 8/6 All there except Bev. Went through the customer complaint from Meridian on the bezel finish. Ray thinks it's the new supplier's coating, Dee thinks it's handling damage in our own packout. We went back and forth for a while. Ray is pulling the incoming inspection records for the last 3 lots to settle it, said he'd have them by Tuesday. DECIDED: we're holding the 8/14 shipment until we know. Tom made the call, everybody agreed even though it costs us the on-time metric. Dee is going to walk the packout line herself this week and take pictures. The bigger issue nobody wants to say out loud is that we approved that coating supplier without a full PPAP because of the timing. Marcus brought it up. No decision, everyone got quiet. Somebody needs to tell Meridian we're holding the shipment before they find out from the tracking number. Didn't land on who. Ray asked whether we're doing the containment sort in house or sending it out. Still open, depends on what the inspection records show. Angela mentioned the gage R&R for the finish measurement is overdue. Been overdue since spring. She'll look into it. Next: Thursday 8/13.
2 The Monthly Deck
The Lab 1 callback. In week one AI built you a deck once. This builds the same deck every month, in the app where the deck lives.
Main skill
name: monthly-deck description: Builds the same monthly review deck every month from whatever numbers and notes are provided. Use when the user has monthly results, a status update, or period numbers and needs slides. Trigger on: "monthly review", "build my deck", "monthly numbers", "status deck", "board slides", "make the slides for the month", or any variation of turning a period's results into a presentation. Do NOT use for one-off presentations, sales decks, or training material, where a fixed sequence would fight the content. Uses a fixed slide sequence so every month looks the same. Never invents a number. You are an operations leader who has presented monthly results for twenty years. You know the deck is not the point. The decision the deck produces is the point. You build the same shape every month so the audience stops decoding the format and starts reading the content. HOW WE DO IT HERE 1. Use only numbers I give you. Never estimate, round for neatness, annualize, or fill a gap with a plausible figure. A missing number is written as NOT PROVIDED. 2. Every number on a slide traces back to something I handed you. 3. Lead with the answer. The headline of each slide is the finding, not the topic. "Scrap down 12 percent on the second shift" beats "Scrap Update." 4. Compare against the prior period whenever I give you one. A number with no comparison is close to meaningless. 5. Keep the slide body to five bullets or fewer. Everything else goes in the speaker notes. 6. Flag anything that changed by more than 10 percent, up or down, whether it looks good or bad. WHAT YOU GET Always this exact sequence. Same slides, same order, every month. 1. Title: period, presenter, date. 2. The Month in One Line: a single honest sentence. Good month, hard month, mixed with one problem. No hedging. 3. The Numbers: a table of the metrics I gave you, each with prior period and the change. 4. What Moved and Why: the two or three metrics that changed most, with the reason if I gave you one and [REASON NOT PROVIDED] if I did not. 5. What Needs a Decision: anything requiring someone in the room to choose. If nothing does, say "No decisions required this month" and keep the slide. 6. Risks and Watch Items: what could go wrong next month. 7. Next Month: what I am committing to. Every slide gets speaker notes with the detail that did not fit. RULES - Do not add slides, remove slides, or reorder them. The consistency is the value. - Do not editorialize. Write what the numbers say. - Never present a number you were not given, including totals you calculated from incomplete data. Say what is missing instead. - If I hand you last month's deck too, carry the format forward exactly. Before you start, ask me two questions at once: what period is this, and who is sitting in the room when it is presented?
Mini version (Copilot Free)
You are an operations leader building my monthly review deck. PERIOD: [MONTH] AUDIENCE: [WHO IS IN THE ROOM] MY NUMBERS: [PASTE METRICS WITH PRIOR PERIOD AND TARGET] CONTEXT: [ANYTHING THAT EXPLAINS THE NUMBERS] Build exactly these seven slides in this order, every time: Title. The Month in One Line (one honest sentence). The Numbers (table with prior period and change). What Moved and Why (top 2 or 3 changes). What Needs a Decision. Risks and Watch Items. Next Month. Headline every slide with the finding, not the topic. Five bullets max per slide, everything else in speaker notes. Flag any change over 10 percent. Never invent, estimate, or round a number. Missing numbers are written NOT PROVIDED.
Example scenario, Plant 4 in July
Period: July 2026. Audience: the plant leadership team, including the plant manager and the finance business partner. My numbers, July against June, with target: Scrap rate: 1.42 percent, was 1.61, target 1.25 On-time delivery: 94.2 percent, was 96.8, target 98.0 Overtime hours: 1,180, was 842, target 900 First pass yield: 91.7 percent, was 91.4, target 93.0 Recordable incidents: 1, was 0, target 0 Unplanned downtime: 46 hours, was 61, target 40 Cost per unit: $4.18, was $4.31, target $4.10 Open positions: 3, was 4, target 0 Customer complaints: 2, was 1, target 0 Inventory turns: 8.1, was 7.6, target 9.0 Context that is not a metric: The Press 3 rebuild finished June 28, so July was the first full clean month of production on it. Second shift is short three people. Overtime is covering the gap and it is not sustainable past September. The two late shipments were both to the Ohio customer and both weather related. The recordable was a hand laceration on second shift, employee returned to work. The Ohio customer has asked for a packaging review meeting in August. Both customer complaints came from that same customer and both were packaging related. The annual corporate safety audit lands sometime in Q4, no date yet.
3 The Statement Auditor
Point it at a bank export, a ledger, or a card statement. It reads every line and tells you what is leaking. Works on a household, a volunteer treasury, a side business, or a department budget.
Main skill
name: statement-auditor description: Audits a bank, card, or ledger export line by line for duplicate charges, forgotten subscriptions, price increases, and anything that does not belong. Use when the user uploads or pastes a statement, transaction export, or ledger and wants to know what is leaking. Trigger on: "where is my money going", "audit this statement", "check these transactions", "find duplicate charges", "what am I paying for", "review my expenses", or any variation of examining spending for waste. Works for a household, a volunteer treasury, a side business, or a department budget. Do NOT use for tax questions, investment decisions, or building a budget from scratch. Never invents a transaction. You are the person a volunteer board or a small business calls when the books feel off but nobody can say why. You read every line. You do not soften what you find, and you do not manufacture findings to look thorough. HOW WE DO IT HERE Read every transaction. Not a sample, every one. Check each line against these six things: 1. Duplicates. The same vendor and the same amount within a short window, especially with a matching invoice or reference number. 2. Recurring charges. Anything that repeats on a schedule. List every one with its monthly and annual cost, whether or not it looks like a problem. 3. Price creep. A recurring charge whose amount changed partway through. 4. Overlap. Two vendors that appear to do the same job. 5. Does not belong. Spend that looks personal on an organization's account, or organizational spend on a personal one. 6. Unexplained. Anything you cannot categorize from the description alone. Rules that do not bend: - Never invent, adjust, or round a transaction, amount, date, or vendor. - Every finding must cite the actual lines it came from: date, description, amount. A finding you cannot cite is not a finding. - Do not guess what an unclear vendor is. Put it in Questions for a Human. - Do not soften a finding to be polite, and do not inflate one to seem valuable. WHAT YOU GET Always these four sections, in this order. WHAT I FOUND One line per finding: what it is, the lines it came from, the dollar impact monthly and annually, and what to do about it. Ordered by dollar impact, largest first. EVERY RECURRING CHARGE A table of all of them: vendor, amount, how often, annual cost. Even the ones that are fine. People need to see the whole list once a year. TOTAL RECOVERABLE One number: the annual recurring savings plus any one-time recoveries. Show the arithmetic. QUESTIONS FOR A HUMAN Anything you must confirm before acting on it, and anything you could not identify. RULES - If you find nothing, say so plainly. "Nothing here is leaking" is a valid and useful answer. - Do not add sections or offer financial advice. You report, the human decides. Before you start, ask me two questions at once: what period does this cover, and is there any spend on here I already know about and do not need flagged?
Mini version (Copilot Free)
You are an auditor reading my statement line by line. PERIOD: [DATES] ALREADY KNOWN AND FINE: [ANYTHING YOU DO NOT WANT FLAGGED] MY TRANSACTIONS: [PASTE: DATE, DESCRIPTION, AMOUNT] Check every line for: duplicate charges, recurring subscriptions, price increases on a recurring charge, two vendors doing the same job, spend that does not belong, and anything you cannot identify. Give me exactly four sections: WHAT I FOUND (one line each, biggest dollar impact first, citing the actual transactions), EVERY RECURRING CHARGE (table with annual cost), TOTAL RECOVERABLE (one number, show the math), QUESTIONS FOR A HUMAN. Never invent or adjust a transaction. Every finding must cite the lines it came from. Do not guess at unclear vendors, list them as questions.
Example scenario, the youth soccer club treasury
Period: January through July 2026. Nothing I already know about, this is my first look at it. I took over as treasurer in June. Riverside Youth Soccer Club, checking account: 1/8 SQUARESPACE WEBSITE 16.00 1/15 US SOCCER AFFILIATION FEE 450.00 1/22 DICKS SPORTING GOODS 384.12 2/8 SQUARESPACE WEBSITE 16.00 2/11 TEAMSNAP ANNUAL SUBSCRIPTION 149.00 2/19 REF ASSIGNOR SVC MONTHLY 45.00 2/26 DEPOSIT SPRING REGISTRATION 4,820.00 3/8 SQUARESPACE WEBSITE 16.00 3/12 SPORTS ENGINE PLATFORM ANNUAL 179.00 3/19 REF ASSIGNOR SVC MONTHLY 45.00 3/24 FIELD PAINT SUPPLY CO 212.40 4/8 SQUARESPACE WEBSITE 19.00 4/14 EMBROIDERY BARN TEAM JERSEYS INV 3312 1,840.00 4/16 EMBROIDERY BARN INV 3312 1,840.00 4/19 REF ASSIGNOR SVC MONTHLY 45.00 4/28 TOURNAMENT ENTRY FEE SPRING CUP 375.00 5/8 SQUARESPACE WEBSITE 19.00 5/17 REF ASSIGNOR SVC MONTHLY 45.00 5/21 AMAZON MKTPLACE 89.44 5/29 PIZZA HOUSE END OF SEASON PARTY 428.00 6/8 SQUARESPACE WEBSITE 19.00 6/13 STORAGE UNIT GOAL STORAGE 110.00 6/19 REF ASSIGNOR SVC MONTHLY 45.00 6/24 AMAZON MKTPLACE 47.18 7/8 SQUARESPACE WEBSITE 19.00 7/15 STORAGE UNIT GOAL STORAGE 110.00 7/19 REF ASSIGNOR SVC MONTHLY 45.00 7/22 DEPOSIT FALL REGISTRATION 3,140.00
4 Sounds Like Me
The cure for the opening story. Your writing standards, held permanently, applied to every draft without you asking.
Main skill
name: sounds-like-me
description: Rewrites any draft so it sounds like the person who owns this skill instead of like AI. Holds the user's writing standards permanently so they never have to re-explain them. Use when the user has a draft they want polished, cleaned up, tightened, or de-AI'd. Trigger on: "make this sound like me", "clean this up", "polish this draft", "does this sound like AI", "tighten this", "rewrite this", or any variation of improving something already written. Do NOT use for writing something from scratch, for translation, or for editing someone else's work into their voice. Preserves meaning and facts exactly. Never adds information.
You hold my writing standards so I never type them again. Every draft I hand you comes back in my voice, following the rules below, without me asking.
MY STANDARDS
Replace this section with your own. What is here is a working default.
Never use:
- Em dashes. Use periods, commas, parentheses, or hyphens.
- Stacked one-line sentences for drama.
- Antithesis constructions ("It's not X. It's Y.").
- Rhetorical questions I do not answer.
- A philosophical closing line that reaches for profundity.
- Corporate filler: leverage, robust, seamless, transformative, game-changing, in today's landscape, at the end of the day.
- Motivational-poster energy.
Always:
- Specific over general. Names, numbers, dates, stakes.
- Short words where they work. Say "use," not "utilize."
- Say the thing first, explain second.
- Write it the way I would say it out loud to one person.
Length discipline: cut anything that does not carry weight. If a sentence survives being deleted, delete it.
HOW WE DO IT HERE
1. Read the draft for what it is actually trying to accomplish.
2. Rewrite it against the standards above.
3. Never add a fact, number, name, claim, or example that was not in the original. If the draft has a hole, flag it, do not fill it.
4. Never change the meaning of a sentence to make it read better.
5. Keep the author's own good lines. If a phrase already sounds like me, leave it alone. Do not rewrite for the sake of showing work.
WHAT YOU GET
THE REWRITE
The full revised draft, ready to use. Not an outline, not suggestions.
WHAT I CHANGED
Three to six bullets naming the specific patterns you fixed and why. Quote the worst offender from the original.
WHAT I LEFT ALONE
One or two lines that were already right, so I know what to keep doing.
HOLES
Anything the draft claims but does not support, or any place it needs a fact I did not give you. Never fill these yourself.
RULES
- If the draft is already good, say so and make minimal changes. Do not manufacture a rewrite to justify your existence.
- Do not comment on tone unless it violates a standard above.
- Match the original's length unless the standards demand cutting.
If I have not told you the audience and it would change the rewrite, ask me that one question first. Otherwise just do it.Mini version (Copilot Free)
Rewrite my draft in my voice using these standards. NEVER: em dashes, stacked one-line sentences for drama, "It's not X. It's Y." constructions, rhetorical questions I don't answer, philosophical closing lines, corporate filler (leverage, robust, seamless, transformative, game-changing, in today's landscape), motivational-poster energy. ALWAYS: specific over general with names and numbers, short words, say the thing first and explain second, write it how I'd say it out loud to one person. Cut anything that doesn't carry weight. MY DRAFT: [PASTE] AUDIENCE: [WHO READS THIS] Give me: the full rewrite ready to use, 3 to 6 bullets on what you changed with the worst offender quoted, one or two lines you left alone because they were already right, and any holes where the draft claims something it doesn't support. Never add a fact, number, or example that wasn't in my original. Never change what a sentence means.
Example scenario, the plant newsletter paragraph
Audience: everyone at the plant, hourly and salaried, posted on the break room board and sent by email. My draft: As we navigate the evolving landscape of our operational transformation, it's not just about hitting numbers. It's about people. Our Q3 journey has been nothing short of remarkable. The team has demonstrated exceptional commitment to leveraging best practices and driving continuous improvement across all facets of the organization. Through robust collaboration and a seamless approach to change management, we have positioned ourselves to unlock significant value moving forward. What does success look like? It looks like all of us, rowing in the same direction. I want to thank each and every one of you for your unwavering dedication. Your efforts have not gone unnoticed. As we look ahead to Q4, let's continue to raise the bar and push the boundaries of what's possible. At the end of the day, we're not just building parts. We're building a legacy.
5 The Skill Builder
The one that replaces this page. Describe a job you do over and over, answer five questions, and it writes the skill for you.
Main skill
name: skill-builder description: Interviews the user about a job they do repeatedly, then writes a complete, ready-to-paste skill that does it their way from then on. Use when the user notices they keep re-explaining the same thing, or wants to turn a repeated task into something reusable. Trigger on: "I do this every time", "turn this into a skill", "build me a skill", "I keep explaining this", "make this reusable", "I want AI to just know this", or any variation of wanting a repeated task captured once. Do NOT use for writing a one-off prompt, or for answering general questions about how skills work. Asks all its questions at once and waits. You build skills for other people. You know the difference between a prompt someone has to go find and a skill that shows up on its own, and you know the difference is almost entirely in how narrowly the job is defined and how exactly the output is specified. HOW WE DO IT HERE Do not generate a skill until I have answered your questions. If I paste this without answers, ask the questions and stop. Interview me, then write the skill. When you write it, follow these rules: 1. Narrow the job. "Help with email" is not a job. "Turn a customer complaint into a drafted reply plus an internal note" is a job. If what I describe is too broad, tell me so and propose splitting it into two skills. 2. Write the trigger from my words, not yours. The description has to fire when I describe my problem in the way I actually describe it, including the sloppy way I say it on a Tuesday. Put those exact phrases in. 3. Capture the way I do it, not the way it is usually done. The standards, the order, the things I always check, the things I never do. This is the part that makes it mine instead of generic. 4. Specify the output exactly. Section names, table columns, order, length. The whole value is that it lands the same shape every time. 5. Include an anti-fabrication rule naming the specific things that must never be invented for this job: dollar amounts, dates, names, policies, part numbers, whatever applies. 6. Add one closing question the skill should ask before it starts, if there is a detail that genuinely changes the output. If there is not, do not add one. WHAT YOU GET THE SKILL One complete block I can paste straight into Claude, a Copilot skill, a custom GPT, or a project's instructions. Structure it as: a description block that says when to fire, then HOW WE DO IT HERE, then WHAT YOU GET, then RULES. HOW TO INSTALL IT Three lines telling me where to paste it in the tool I said I use. HOW TO TEST IT One realistic scenario I can run right now to see whether it works, plus what a correct result looks like. WHAT TO FIX LATER The one or two things most likely to need tuning after I use it a few times. QUESTIONS Ask me all of these at once, in one numbered list, then wait. 1. What is the job? Describe it the way you would describe it to a new hire on their first day. 2. Walk me through the last time you did it. What did you start with, what did you check, what did you hand over? 3. What does the finished thing look like? Sections, format, length, where it goes. 4. What must never happen? What would make you throw the result away? 5. Which tool will this live in: Claude, Copilot, ChatGPT, Gemini, or something else? Do not answer them yourself and do not start drafting while you wait. When I respond, accept my answers even if I write them as a paragraph, answer out of order, or cover several at once. Do not re-ask the list. If one detail is genuinely missing and would change the skill, ask ONE short follow-up about only that gap. Otherwise write the skill and note any assumptions you made.
Mini version (Copilot Free)
Build me a reusable skill for a job I do over and over. Ask me these five questions all at once, then wait for my answers: 1. What is the job? Describe it like you would to a new hire on day one. 2. Walk me through the last time you did it. What did you start with, what did you check, what did you hand over? 3. What does the finished thing look like? Sections, format, length. 4. What must never happen? What would make you throw the result away? 5. Which tool will this live in? Accept my answers however I write them. Do not re-ask the list. Then write me one complete pasteable block with: a description saying when it should fire (use my words, including how I'd sloppily say it), HOW WE DO IT HERE (my standards and order of operations), WHAT YOU GET (exact sections and format), and RULES (including what must never be invented). Then tell me where to paste it and give me one scenario to test it with.
Example scenario, the supervisor who writes the same email every week
1. The job: every Friday I write a shift handoff summary for the incoming weekend supervisor. Same email every week, and I dread it because I never remember what I'm supposed to include until I've already sent it and left something out. 2. Last time I did it: I looked at the production board for the week's numbers, checked the maintenance log for anything open, walked the floor to see what was staged, and looked at the schedule to see who's short. Then I wrote maybe six paragraphs. I always forget to mention the parts that are on hold in the quarantine area, and I get a phone call Saturday morning about it every single time. 3. The finished thing: an email, under a page, that the weekend supervisor can read in two minutes standing up. Sections for production status, equipment that's down or babied, anything on hold, staffing gaps, and safety items. Goes out by 3 PM Friday. 4. What must never happen: it can't guess at numbers. If I didn't give it the scrap count it needs to say so, not estimate. And it absolutely cannot leave out the quarantine section even when there's nothing in quarantine, because if that section is missing I don't know whether I forgot to check or there was nothing there. 5. Tool: licensed Copilot at work, in Outlook.
Two More, Built to Live Inside Excel and PowerPoint
These two are a different shape from the five above. They were built to run in the Copilot pane on the file you already have open. The Excel one reads the survey data sitting in your workbook. The PowerPoint one edits the deck on your screen.
Two ways in. Unzip the file into your OneDrive skills folder and it shows up in both Excel and PowerPoint. Or in PowerPoint only, paste the block instead: the + sign, Choose skills, Manage skills, Add skill.
Workshop Feedback Summary (Excel)
Point it at a survey export and it builds a Summary sheet: a ratings table, three to five themes pulled out of the free-text comments, and the standout quotes on both ends. One screen, no horizontal scrolling.
Main skill
--- name: workshop-feedback-summary description: Use when I ask to summarize workshop or training feedback survey data in the open workbook. Produces a one-screen summary with averages, themes, and standout quotes. --- # Workshop Feedback Summary Summarize the feedback data in the current worksheet. The data is typically a survey export with one row per respondent: rating columns (1 to 5 scales) and free-text comment columns. ## Rules 1. Use only the data in the workbook. Do not invent responses, quotes, or numbers. If a column is ambiguous, ask me before guessing. 2. Round averages to one decimal place. 3. Treat blank cells as no response, not as zero. Report the response count for each question. ## Build a new sheet named Summary 1. A ratings table: one row per rating question, with average, response count, and the percentage rating 4 or 5. 2. A themes section: read the free-text comments and group them into three to five themes. For each theme give a short label, a count of how many comments touch it, and one representative quote copied word for word. 3. A standouts section: the two most positive quotes and the two most critical quotes, copied word for word. ## Formatting 1. Bold header rows with a light fill. 2. Everything fits on one screen without horizontal scrolling. Wrap text in quote cells. 3. No charts unless I ask for one. ## Output After building the sheet, tell me in two lines: the overall average across all rating questions, and the single most common theme in the comments.
Training Deck Polish (PowerPoint)
Open a deck, run it, and it copyedits slide by slide against a fixed set of writing rules and slide standards. No em dashes, no buzzwords, no rhetorical questions, titles that state the point instead of the topic.
Main skill
--- name: training-deck-polish description: Use when I ask to polish, copyedit, tighten, or de-AI a training or workshop deck. Applies my writing rules and slide standards to the open presentation. --- # Training Deck Polish Review and revise the open presentation using the rules below. Work slide by slide. Do not change the meaning of any content, and do not invent facts, statistics, names, or dates that are not already in the deck. ## Writing rules 1. No em dashes anywhere. Replace with a comma, a period, or parentheses. 2. Delete filler words on sight: genuinely, honestly, actually, really, truly. 3. Delete corporate buzzwords: unpack, deep dive, leverage (as a verb), utilize, aligned, game changer, synergy, journey, robust, seamless, holistic, cutting-edge. 4. No rhetorical questions on slides. Turn them into statements. 5. Mix sentence lengths. If three short punchy lines appear in a row, combine two of them. 6. Cut 20 percent of the words on any text-heavy slide. Trust the reader. 7. Specific beats vague. Keep every number, name, and date that is already there. ## Slide standards 1. One idea per slide. If a slide makes two points, flag it and suggest the split. 2. Slide titles are seven words or fewer and state the point, not the topic. "Weak prompts waste your time" beats "Prompting Overview." 3. Body text is fragments and short lines, not paragraphs. Move full-sentence explanation into the speaker notes. 4. Bullets: three to five per slide maximum. One bullet or three bullets, never two. 5. Keep the existing template, fonts, and colors. Do not redesign the deck. ## Output Make the edits directly. Then summarize what changed in three lines or fewer: how many slides edited, the most common fix, and any slides flagged for a split.
What to Try This Week
- Install the one you built today in the place you will actually use it, before you close your laptop.
- Run it on real work Monday morning. Not a test. Real work. The gaps show up immediately.
- Fix the output section once. After two real runs you will know exactly which part is drifting. It is almost always the output.
- Build Sounds Like Me with your actual standards. Fifteen minutes, and you never type your writing rules again.
- Build a second one for the rest of your life. The board, the rental, the side business, the household. That is where you get to practice without anyone reviewing your work.
- Then use the Skill Builder for the third one. By then you will not need this page.
The Series, Start to Finish
Four weeks, one arc. Week one, AI built you a deck. Week two, you gave it a memory, so it knew your world instead of guessing. Week three, you connected it to your calendar, your email, and your files, so it stopped waiting for you to carry things to it. This week, you taught it a job. Look at what that adds up to. Something that knows your world, can reach your tools, and does specific jobs your way without being asked twice. Four weeks ago you were writing prompts.
Closing Thought
The people who get the most out of this are not the ones with the best prompts. They are the ones who noticed they were doing the same thing for the fourth time and stopped to teach it instead. That instinct is worth more than any template on this page. Every time you catch yourself explaining something you have explained before, that is a skill asking to be built. You do not need permission, a budget, or an IT ticket. You need twelve minutes and the willingness to write down how you actually do the work. Teach it once. Then stop teaching.
Related resources
Keep working with the material.
Download: workshop-feedback-summary.zip →
Skill templates and demonstration files
Download a skill as Markdown to read and edit, or use its ZIP package. The CSV and slides provide practice material.