There is a calculator Steph has wanted for about three years, and it does not exist.
Four boxes and an answer. What a groom is worth. How many people did not turn up last month. What she collected off them. How many of those slots she refilled. Then one number telling her what the empty chair cost her, this month and across a year.
She has looked. Nothing for sale fits, which is the first clue that this is a tool you build without coding rather than buy. What comes back wants her email address before it will show her the total, or is built for a dental surgery with fields for chair-hour cost and an insurance rebate. Neither knows what a de-matting surcharge is, and neither knows her prices.
Nobody sells it because it is not hard. A calculator like this is only worth opening if the numbers inside it are hers, and nobody can sell her her own numbers.
The direct answer. You build tools without coding by describing the job in plain English to Claude Code, which writes a small file on your own computer, runs it, and prints the answer. Two things make it yours rather than somebody else's product: the numbers live in a plain text file you edit, and when the answer is wrong you name the part that is wrong and it changes. Not an app, not a website. A small thing in a folder that answers one question.
Take Steph, two chairs and one bath in Sydney's Inner West. She is the worked example this series follows, and her figures are the shape a real salon's would be rather than any real salon's.
What building tools without coding actually means
Two assumptions: Claude Code installed and signed in, and a memory holding your prices and policies. If neither means anything yet, start with what Claude Code actually is. What all of this costs is a post of its own.
Keep the scope small and it stays honest. Nothing here gets hosted, nobody logs into it, and it has no address on the internet. It is a file on your laptop that answers one question and writes the answer down, so you can read it again in March.
Everything you build has two halves. A plain text file holding your numbers, which you write and own, so a price rise means editing six lines rather than rebuilding anything. Then the tool that reads it and does the arithmetic. All of it happens in the terminal, a window you have by now used for duller things than this.
Tool one: the no-show cost calculator
Build this one first, because you can catch it lying: every figure in it is a multiplication you can do on your phone. It runs the revenue side of the framework for what a no-show costs on your numbers, whenever you want it.
What to ask for
Your numbers file first, saved as no-show-numbers.md, because that is the part that matters:
Then the tool:
What you get back
| Service | Price | Missed | Recovered | Refilled | Net a month |
|---|---|---|---|---|---|
| Full groom, medium | $95 | 5 | $95 | $95 | $285 |
| Nail trim | $20 | 2 | $0 | $0 | $40 |
| Total | 7 | $95 | $95 | $325 |
Five missed full grooms at $95 is $475 of work that did not happen. Her policy charges 50% inside 24 hours' notice, so the two late cancellations returned $47.50 each, being $95. She refilled one slot, another $95. That leaves $285 on the grooms, and the nail trims return nothing, so $40 on top.
$325 a month. $3,900 a year.
How to check it is right
One row, one calculator, half a minute.
Five times 95 is 475. Two times 47.50 is 95. One times 95 is 95. Then 475 less 95 less 95 is 285.
If the row says anything else, there is one likely culprit. Charge the policy percentage on all five misses rather than the two late cancellations and you recover $237.50, which turns $285 into $142.50 and makes your cancellation policy look twice as good as it is.
Every skill in our library carries a check step for this reason: a table with dollar signs in it sounds more confident than it has any right to.
How to change it when it is wrong
Steph's first run counted a cancellation given three days out as a late one, because her file said "cancellations" and she had meant something narrower. She said so: inside 24 hours only.
That fix needed no developer and no understanding of what the tool is made of. She read an answer, found the sentence that was wrong, and named that sentence. That is the step that makes you the owner of the tool rather than its user, and it is the one most people skip, because a wrong answer feels like a reason to stop rather than the second half of the job.
One habit makes it stick. Ask for the correction to go into the tool rather than into this one answer, then run it again from cold. If the second run gives the corrected figure unprompted, it landed. If not, it patched the output and left the tool alone, and you say so.
Tool two: the matting time and price estimator
This is the one she needs standing up. Moose the Groodle arrives matted every time, and her policy is that if the coat needs clipping short she says so before she starts, not after. Which means a number, out loud, at the desk, with the owner watching.
Nothing can tell her how long a matted coat takes. Not a chart, not a forum, not an AI.
Only her own recorded times can, so the first step is a fortnight with a card by the table: size band, coat condition on a scale of one to four she defines in a line each, and the real minutes from on the table to off it. Not the slot she booked. Fourteen days is not many dogs, which is why the refusal rule matters more than anything else in the file.
What to ask for
That last rule is not decoration. A tool that scores coats is one careless sentence from sounding like it has examined a dog, which is where AI must never go in pet care.
What you get back
For medium dogs at condition 1 she had three measured grooms: 70, 75, and 80 minutes. Average 75. At condition 3 she had 105, 120, and 135. Average 120.
So a badly matted medium coat costs her 45 extra minutes of her own bench. Her price list books a medium at $95 for 75 minutes, which is $1.2667 a minute, or $76 an hour. Those 45 minutes are worth $57. Her current de-matting surcharge is $30.
It refused two rows and said so plainly: small at condition 4 had one measured groom, large at condition 2 had two.
The $57 is not the tool telling her to charge $57. It is 45 minutes of her own bench at her own rate, sitting next to the $30 she charges today, and the $27 gap is hers to decide about. Costing a bench hour properly is its own piece of arithmetic; this is a rough version, per coat, and better than picking a number with a matted Groodle in front of her.
How to check it is right
Two divisions. 105 plus 120 plus 135 is 360, over three is 120. And 95 divided by 75 is 1.2667, times 45 is 57. Both by hand, on the row you care about.
Then the check that catches the failure that actually happens: count the measured grooms in the table it printed against the number of lines in your times file. If they do not match, it has dropped the rows it could not read, and an average of the rows it liked is not an average.
How to change it when it is wrong
Run one gave her a medium condition 2 figure off two grooms, because two is nearly three and it wanted to be useful. She told it that fewer than three refuses, no exceptions, and that refusals print as prominently as results. Run two refused from cold, unreminded.
The second fix was not the tool. Three of her fours should have been threes, because she had been scoring how annoying the coat was rather than how matted it was. She rewrote the four lines defining her scale, re-scored the fortnight, and ran it again. Same tool, different answer, and the tool was never the problem.
A tool you build without coding is only as good as your assumptions
Both of these are arithmetic on numbers you typed in. That is their strength and the only real risk here.
If your no-show count is a guess, the $3,900 is a guess with four figures on it. On the page it looks far more solid than it felt in your head, and the tool cannot tell you otherwise, because it has no idea where a number came from.
So write the provenance into the numbers file, as the template above does. Count what you can count, label the rest as guessed in writing, and when a guessed figure turns out to be carrying the answer, that is your instruction for the next fortnight.
What these hold, and what they must never hold
Both tools hold your numbers. Neither holds a client's details, and neither needs to. The calculator wants a count, not a list of names. The estimator wants a size band and some minutes; a dog's name helps you find a row again, and an owner's name, phone number, or address has no job in either file.
That is not squeamishness. A file you built on a Tuesday afternoon has no access rules and no record of who opened it, and it will still be in that folder when your laptop goes in for repair. Two lines cover it:
Never paste into an AI chat:
The full list, and the law behind it, is worth reading once properly. The version that applies here is one sentence: your prices are yours to hand around, and your clients' details are not.
Do not loosen the permission setting it came with, the one that asks you first. A job that reads two text files and writes one report needs nothing more relaxed than that, and what to allow and what never to deserves its own reading.
Some of what you build gets thrown away
Steph built four things that month. Two are in this post. The third was a rota tool that turned out to be a worse version of the calendar she already had. The fourth was going to work out which services earned her most per hour, and she abandoned it on realising she had no recorded times to feed it.
That abandonment became the fortnight of measuring that made the estimator possible.
Two out of four, for one afternoon. That is a good rate.
A tool you paid for has to be worth keeping, because you paid for it, so you carry on using something that half works. A tool you described in four minutes can be deleted in one with no feeling about it. Building the wrong one is how you find out what the right one is.
When a tool becomes a skill
The first time, you describe the tool. The second time you want it, you describe it again, slightly differently, and get a slightly different answer. That is the moment to write it down instead.
A tool you reach for repeatedly belongs in a skill, which is the same instruction in a file, so it behaves identically every time and you never explain it again. The estimator is the candidate, because Steph reaches for it at 8:40am with a coat in front of her. Moving it across is mostly copy and paste: the prompt above already has what it needs before it starts, its steps, its refusals, and its check, which are the four headings a skill file wants. Twelve written out in full sit in the skill library, and price-check in there refuses the same way: too few figures gathered, no answer, and it names how many more it needs.
The calculator is the opposite case. You run it twice a year, when you are deciding whether to change a policy, so leave it as something you describe when you need it, and let the Monday morning routine have the jobs that genuinely repeat.
What this actually reclaims for you
The calculator took about 25 minutes, including the runs that were wrong. The estimator took 20 minutes, plus a fortnight of writing three things on a card by the table.
Neither reclaimed an hour, and that would be a strange claim about something you run twice a year.
What the calculator produced was $325 a month and $3,900 a year, on her own counts at her own prices. That is not $3,900 reclaimed. It is $3,900 she can now see, which is the whole distance between a cancellation policy she keeps meaning to enforce and one she enforces on Tuesday. The estimator produced $57 against a $30 surcharge, which is a $27 conversation she can have before the clippers go on rather than after.
The thing she actually got back is smaller than either figure and worth more than both. For three years there was a calculator she wanted and could not buy. Now, when she wants one, she builds it, badly, on a Tuesday, and fixes it on the Wednesday.
Next in the ladder is a full week of this running at once, and then asking your own business a question and getting an answer out of your own exported data.