Introduction — The Column That Had Been Wrong for a While

Two hundred and sixty-six calculations, none of them ever checked, and a customer's accountant found the first one.

Rae had built a bookkeeping workbook for small service businesses. Two hundred and fourteen formulas, thirty-one validation rules, twelve named ranges and nine conditional formatting rules — two hundred and sixty-six calculations in all, put together across about seventy hours over five weeks. It looked professional. The dashboard was clean. It had a category summary, a quarterly view, a running profit figure and a chart that updated itself.

It sold. Not spectacularly, but steadily, and the reviews were kind. People said it was clear. People said it was exactly what they had been looking for.

In month six, one of them sent an email.

The email was not angry, which is part of why it took so long to arrive. It said that the quarterly totals column had "been wrong for a while," that they had noticed because their accountant queried a figure, and that they had gone back through it and thought the problem started somewhere around the second quarter. They wanted to know whether it was them or the sheet.

⚠ It was the sheet.

---

What was actually wrong

The quarterly summary used a formula that summed a column conditionally on the date falling inside a range. Rae had written it, looked at it, seen a plausible number appear underneath it, and moved on.

In Rae's own copy, with Rae's own sample data, typed on Rae's own machine, it was correct. It was correct because every date in the sample had been entered by Rae, in one format, on one system, in one application.

The customer's dates had been typed by a person who wrote dates the other way round. Some of them had been pasted in from a bank export. A handful had arrived as text rather than as dates and looked identical on screen.

The formula did not complain. It did not show an error. It did not turn red. It added up the rows it recognised as dates, ignored the ones it did not, and printed a total that was smaller than the truth by an amount neither Rae nor the customer could see.

⚠ That is the entire problem with this business, in one sentence: the failure mode of a spreadsheet is a confident wrong answer.

A broken image looks broken. A broken document looks broken. A broken layout looks broken. A broken spreadsheet looks like a spreadsheet, and the number in the corner is the one the customer is about to make a decision with.

---

The number Rae had never produced

After the email, Rae did something that took an afternoon and should have taken an afternoon five months earlier.

Rae sat down with a calculator, a sheet of paper and the workbook, and went through the formulas one at a time. For each one: what should this return, given this specific input, worked out by hand? Then: what does it return?

Not "does it look sensible." What is the correct answer, and does it produce it?

Two hundred and sixty-six calculations. It took most of two days spread across a week. And it was not one pass, because a formula that is correct with tidy data is not necessarily correct with the data a real person produces. So each formula was checked against five inputs that Rae came to think of as the hostile five:

• an empty cell where a number was expected
• text typed into a number column — a currency symbol, a word, a space
• a date entered in a different format from the one Rae uses
• a negative number
• the same sheet opened in the other application Rae had claimed to support

⚠ Thirty-one of the two hundred and sixty-six produced a wrong or misleading result under at least one of those five.

Not thirty-one that showed an error. Thirty-one that returned something — a number, a blank, a total — that was not the right answer and did not announce itself.

Some were small. A count that quietly excluded rows. A percentage that divided by a total that had itself been filtered. Some were not small at all. The quarterly summary was one of them, and so were three of the four figures on the dashboard, because the dashboard was built on top of the summary and inherited everything wrong with it.

⚠ The dashboard is the most dangerous sheet in the workbook, because it is the one the customer looks at and the one furthest from the data.

---

Hand-check coverage

Before that week, if you had asked Rae how many formulas in the workbook had been verified against an answer worked out independently, the honest number was zero.

Not zero because Rae was careless. Zero because there had never been a moment in the process where that was a step. Rae had built formulas, looked at the result, thought yes, that's about right, and moved on. Everybody does. The result appears instantly, it is formatted, it sits in a cell that looks authoritative, and the mind accepts it.

⚠ Looking at the output of a formula is not checking it. It is being told the answer by the thing you are trying to check.

So Rae started keeping a number, and it is the number this book is organised around.

Hand-check coverage: the share of the calculations in a template whose output has been compared against an answer worked out independently, under every one of the hostile inputs, in every environment you claim to support.

Rae's coverage in month five was zero per cent. By month eleven it was one hundred per cent — not because Rae had checked two hundred and sixty-six formulas, but because Rae had rebuilt the workbook to have sixty-eight.

---

The rebuild

The rebuilt workbook did the same job. It supported the same decision. It had one data table instead of four, a set of views built from that table, a much simpler quarterly summary, and a dashboard with five figures instead of eleven.

Sixty-eight calculations. It took about twenty-six hours instead of seventy. Every one of them had a written test case with an expected answer, recorded in a file that lived alongside the master.

And two other things changed, both of which are rules in this book.

Rae stopped claiming to support "any spreadsheet application" and named two, having actually opened the file in both and run the hand-check in each. The listing said so in the first three lines.

And Rae stopped shipping a workbook where the customer could type into a formula cell without noticing. The input areas were obvious and unprotected. Everything else was locked, and the sheet said so.

⚠ The version with sixty-eight calculations sold better than the version with two hundred and sixty-six, and generated roughly a third of the support.

---

Revenue per verified formula

There is an arithmetic that follows from all of this, and it is the second thing this book is organised around.

If a formula has not been hand-checked, it is not an asset. It might be right. It is untested, which means you do not know, which means it can produce a wrong number in a customer's hands at a moment you will not witness. In the arithmetic of this business it is worth nothing, and it is worse than nothing, because it carries a liability and a support cost and it has to be re-checked every time the application updates.

⚠ So the denominator is not the number of formulas you wrote. It is the number you have verified.

Rae's first workbook, on that basis, had a denominator of zero, which is not a number you can divide by. That is the correct answer. It is a product whose economic value per unit of verified work was undefined, because no verified work had been done.

The rebuilt workbook had sixty-eight verified formulas and produced more revenue than the original. That is the whole argument of this book compressed into two sentences.

---

What this book is

Thirty-seven chapters, sixty-eight resources and a hundred and forty-eight prompts, built on four rules:

Rule 1 — A broken spreadsheet does not look broken; it gives you a number. Wrong answers arrive with the same confidence as right ones, from a cell that looks exactly like every other cell. Chapter 5.

Rule 2 — They will type into it, and typing is how it breaks. The product invites the customer to edit the thing that makes it work. Chapter 6.

Rule 3 — It has to run somewhere you have never been. Different applications, versions, devices and locales — and a locale can change the correct answer without changing anything visible. Chapter 7.

Rule 4 — You never see the data it runs on. It is their money, they will not send it to you, and you cannot reproduce a fault you cannot see. Chapter 8.

The economic unit is revenue per verified formula. The leading indicator is hand-check coverage, which you must produce yourself, which takes an afternoon, and which moves a season ahead of refunds, fault reports and reviews.

---

What this book will not do

It will not name an application, a marketplace or a function. It will not tell you what a formula should be, because a formula you did not verify is not yours regardless of who wrote it. It will not tell you what the law requires where you live, what your marketplace's terms currently say, or what your tax obligations are. And it will not tell you that two hundred and sixty-six formulas is a better product than sixty-eight.

⚠ What it will do is put a step into your process that almost nobody in this trade has: before you publish, you find out whether the numbers are right.

Rae's summary of the whole thing, given some two years later, was this:

⚠ "For five months I had a product whose central feature was arithmetic, and I had never once checked the arithmetic. It took two days to find out that thirty-one things were wrong. The customer who told me had already sent the figures to their accountant."

---

©2026 James Henderson / https://localhandyman.work

Chapter 1 — Why Spreadsheet Templates Work as a Side Hustle

You sell a structure and a promise that the arithmetic works. Only one of those is earned.

1.1 What you are actually selling

You are not selling a spreadsheet.

A spreadsheet is free. The application is free or close to it, blank files are unlimited, and anybody who wanted a grid of cells already has one. If the grid were the product, this business would not exist.

⚠ What you are selling is somebody else's decision about how to structure a problem.

The customer knows they should be tracking their expenses. They know roughly what they want to see at the end of the month. What they do not have is the twenty or forty hours of thinking that turns "I should track my expenses" into a set of columns that will still make sense in November, a category list that does not need rewriting in March, a summary that answers the question they will actually ask, and a set of calculations that produce the right number.

They are buying an answer to the question how should this be arranged? — and, underneath it, a promise that the arithmetic works.

⚠ The second half of that is the part sellers forget, and it is the part this book is about.

A layout can be a matter of taste. A total cannot. If your template says a business made a certain amount last quarter, that is either true or false, and the customer cannot tell which by looking.

---

1.2 Why the free templates have not ended this business

Every major spreadsheet application ships with a gallery of free templates. There are thousands more available at no cost from banks, accounting firms, government agencies and personal-finance writers. This has been true for twenty years and the market for paid templates has not disappeared.

There are four reasons, and it is worth being precise about them, because two of them are things you must earn rather than assume.

Free templates are usually general. They cover a broad case in order to be useful to everybody, which means they fit nobody's actual situation without work. A template built for one job — a mobile dog groomer's mileage and materials, a market-stall trader's stock and takings, a school department's budget against a fixed allocation — fits without the customer having to redesign anything.

Free templates rarely come with instructions. They come with an interface and the assumption that you already understand it. A paid template that includes a setup sheet, a walkthrough and an answer to the four questions everybody asks is a different product.

Free templates are not supported. Nobody will answer an email about the one that came with the application.

⚠ Free templates are usually not verified either — and neither are most paid ones. This is the fourth reason and it is the only one that is entirely yours to earn. The default state of a template on the open market is nobody has checked whether the numbers are right. If you are the seller who has, you are selling something materially different, and you can say so.

You cannot say so unless you have done it. Chapter 21.

---

1.3 The moment the customer buys in

Somebody buys a spreadsheet template at a specific kind of moment, and it is almost never a calm one.

They have been meaning to do this for months. Something has now forced it: a tax deadline, an accountant's request, a business that has grown past what memory can hold, a new year, a new job, a bill that arrived and could not be explained. They have half an hour and they want the problem to be solved by the end of it.

⚠ They are buying the elimination of a decision, under time pressure, on the strength of a screenshot.

Three things follow from that.

They will not read a long description. They will look at a picture of the dashboard, decide whether it looks like the thing they had in their head, and buy or leave.

They will open it immediately and expect to understand it in about two minutes. If the first sheet is a wall of instructions or a blank grid with no indication of where to start, a substantial fraction stop there and do not come back.

⚠ And they will trust the numbers completely. Nobody buys a template and then independently verifies the arithmetic. The entire proposition is somebody who knows how to do this has done it for me. That trust is the product, and Rule 1 exists because it is so easy to betray it without knowing.

---

1.4 What a template costs to build, honestly

Rae's first workbook took about seventy hours. That is not an unusual figure and it is not padding.

Roughly, and in the order the hours actually go:

Deciding what it does. Five to ten hours if done properly, and most sellers spend about twenty minutes. This is the requirements work in Chapter 11, and skipping it is what produces a template that grows a new sheet every time you think of something.

The data structure. Three to eight hours. What are the columns, what is one row, what does the customer type and what is derived. Getting this wrong is the single most expensive error available, because everything else is built on top of it.

Formulas. Fifteen to forty hours depending on ambition, and this is where the count runs away from you. Every summary view adds formulas. Every dashboard figure adds formulas. Every conditional format is a rule that has to be right.

Validation and protection. Three to eight hours, usually skipped, and Chapter 6 is about why that is expensive.

Sample data. Two to five hours if you do it properly, which means data shaped like real data rather than tidy invented rows.

Instructions and the setup sheet. Three to eight hours. Nobody enjoys this. It reduces support more than anything else you can do.

The listing, the screenshots and the description. Two to six hours.

⚠ And the hand-check: an afternoon to several days, depending entirely on how many calculations you allowed yourself to build. This is the line item almost nobody has, and it is the reason the calculation count matters more than any other number in the build.

Resource 39 costs your hours out properly. Resource 30 counts your calculations before you commit to them.

---

1.5 What it costs to keep

A template is not finished when it is published. It has a maintenance tail, and the tail is proportional to the calculation count.

Applications update. A function changes behaviour, a feature is deprecated, a default alters. This is rare enough to feel unlikely and frequent enough to catch you at least annually. Every update means a re-check, and the re-check is the hand-check again.

Fault reports arrive. Some are the customer's fault, some are yours, and you cannot tell which without reproducing it (Chapter 33).

Questions arrive. In Rae's shop, questions ran roughly: how do I add a row twenty-seven per cent, it will not work in my application twenty-one per cent, can I change the categories nineteen per cent, the total looks wrong nine per cent, and everything else twenty-four per cent.

⚠ The nine per cent is the one that matters, and it is the smallest category. Most people who find a wrong number assume they did something. They do not write. Rae's month-six email arrived after five months precisely because the customer assumed for most of that period that the fault was theirs.

The catalogue needs reviewing. Quarterly, honestly, against Resource 47 and Resource 48.

---

1.6 The four rules, introduced

The four rules are properties of the product, not opinions about the market. You cannot arrange your way out of them.

Rule 1 — A broken spreadsheet does not look broken; it gives you a number. There is no visual signal for a wrong answer. Chapter 5.

Rule 2 — They will type into it, and typing is how it breaks. You are shipping a product whose interface invites the customer to edit its machinery. Chapter 6.

Rule 3 — It has to run somewhere you have never been. Applications, versions, devices, locales. A comma and a semicolon can be the difference between a formula and an error, and a date format can be the difference between a right answer and a wrong one. Chapter 7.

Rule 4 — You never see the data it runs on. It is their money. They will not send it. You must design, test and support blind. Chapter 8.

Resource 3 is these four on one page. Print it and put it where you build.

---

1.7 Who this business suits

⚠ People who enjoy being wrong on purpose. The hand-check is an exercise in trying to make your own work fail. If you find that satisfying rather than demoralising, this trade fits you unusually well. ⚠ Resource 1.

People who think in structures. The valuable part of the work is deciding what one row represents. If that sentence makes sense to you, you have the main skill.

People who can write plainly for somebody less confident than themselves. Instructions are a product component here (Chapter 23), not an afterthought.

People with domain knowledge. A trade, a profession, a hobby, a school subject, a sector you worked in. Knowing what a small building contractor actually needs to track is worth more than knowing eleven ways to write a lookup.

People who can leave things out. The best commercial move available to you is deletion, and it is the hardest.

---

1.8 Who it does not

People who want passive income. The catalogue needs maintaining, the applications change, and the fault reports do not stop.

People who enjoy building clever things. Cleverness has a strictly negative value here. Every advanced construction is a formula the customer cannot repair, a function that may not exist in their application, and another line on the hand-check. Chapter 20.

People uncomfortable saying "I do not know." You will be asked accounting questions, tax questions and financial questions. The answer is a referral, always. Chapter 15.

⚠ People who cannot bear to be responsible for a number. If a wrong total in somebody's business would keep you awake, be honest with yourself now — because it will happen at least once, and the correct response is a correction and a notice, not a collapse.

---

1.9 What the first year usually looks like

Not a forecast. A shape.

Months one to three. You choose a customer and a job, write requirements, build a structure and start on formulas. You publish something at the end of this, usually one template. It sells a handful of copies and most of those are people who found it by accident.

Months four to six. Questions arrive. They are not the questions you expected. You write the instruction sheet you should have written first. Somewhere in here, if you are doing this properly, you run your first hand-check and find things.

⚠ Almost everybody finds things. Rae found thirty-one. The number is not a sign that you are bad at this; the absence of the number is.

Months seven to nine. You either rebuild or you keep patching. The rebuild feels like going backwards and is the correct choice. You start a second template, chosen from evidence rather than enthusiasm (Chapter 28).

Months ten to twelve. A bundle. A first annual edition if your job has a yearly cycle. Reviews start doing the selling. Your calculation count is lower than it was and your revenue is higher.

⚠ The year-one milestone worth aiming at is not a revenue figure. It is: every calculation in everything you sell has a written expected answer, and you have checked it in every environment you claim.

---

1.10 Chapter summary

You are selling a structure and a promise that the arithmetic works. Free templates have not killed this trade because they are general, undocumented, unsupported and unverified — and the fourth is the one you can earn. The customer buys under time pressure, on a screenshot, and trusts the numbers absolutely. A template takes tens of hours and carries a maintenance tail proportional to its calculation count. The four rules are properties of the product. The trade suits people who enjoy trying to break their own work and can leave things out.

---

AI prompts for this chapter

Prompt 1 — The structure behind the request

"I am considering building a spreadsheet template for [describe the customer and the job in one sentence]. Do not design the spreadsheet. Instead, tell me what decision this customer is actually trying to make, what question they ask at the end of the month, and what one row of data would have to represent for that question to be answerable. Then list the three ways a template for this job most commonly gets its structure wrong. Ask me anything you need before answering."

Prompt 2 — The free competitor

"Free spreadsheet templates for [job] are widely available from application galleries, banks and finance writers. Argue, as strongly as you can, that nobody should pay for a template for this job. Then tell me the smallest number of things a paid version would have to do differently to be worth buying, and mark which of those I would have to earn rather than simply claim."

Prompt 3 — Hour estimate before commitment

"Estimate the build hours for a spreadsheet template that does [describe scope], broken into: deciding what it does, data structure, formulas, validation and protection, sample data, instructions, listing, and hand-checking every calculation against an independently worked answer. Assume the hand-check tests each calculation against an empty cell, text in a number column, a date in another format, a negative number, and a second application. Then tell me what to remove to halve the total."

Prompt 4 — Fit and temperament

"I am deciding whether to sell spreadsheet templates as a side hustle. Interview me with ten questions covering: my tolerance for being responsible for a wrong number in somebody's business, my willingness to spend a full day trying to break my own work, my ability to write instructions for somebody less confident than me, whether I have domain knowledge in a specific trade or subject, and whether I can delete work I am proud of. Ask them one at a time and tell me at the end which answers concern you."

---

⚠ AI checkpoint for this chapter

One — did it treat this as a layout business? It is an arithmetic business wearing a layout.

Two — did it assume the customer checks the numbers? They never do. That is the whole proposition.

Three — did it count validation and formatting as calculations? They are, and they can be wrong.

Four — did it estimate the hand-check as a token hour? Across five inputs and two applications it is not.

---

Do This Now

1. Write down the decision your customer is trying to make, in one sentence.

2. Count every calculation your current template contains, including validation and formatting rules.

3. Pick one formula and work out on paper what it should return for a specific input.

4. Enter that input and compare.

5. Record the result, dated, in a file that lives beside the master.

---

©2026 James Henderson / https://localhandyman.work