For Claude Cowork and Claude Code · even if you’re not a programmer

Proves. Asks first. Remembers. Doesn’t go in circles.

Your AI says “done.” This makes it show the receipt.

Claims Need Receipts is a ready-made set of rules and check scripts for Claude. Every claim about time, numbers and status needs a measured receipt, and “it can’t be done” is only allowed after an actual attempt. Before any step it can’t undo, it waits for your word. It keeps what you agreed on in plain files on your own computer, not in an AI memory you can’t see. And it reads what already failed, so it doesn’t send you down the same dead end twice.

Other AI kits make the AI do more. This one checks what it already claims.

Pay once — $199. No subscription, no monthly fees.
See what this costs you Buy once — $199
  • One-time payment, no monthly fees, tax included
  • 14-day money-back guarantee
  • Your AI walks you through the setup
Sound familiar?

Claude says “done,” and it isn’t. The same mistakes, over and over.

From a long piece of work, AI remembers only fragments. So it goes back to things you solved long ago – or the other way round: it forgets what it was supposed to work on, and doesn’t do it. Every wrong turn like that costs you extra hours of work.

Done.

An hour later you find out it isn’t. It sounded exactly like finished work.

Could you remind me of the context?

A new conversation, and only fragments of what you agreed on are left. You explain it all over again.

Try this next.

And then the next thing. Each tip ignores the one before it — and by evening you’re right where you were in the morning.

I checked it.

It didn’t. It sounded convincing, but there’s no proof anywhere.

Now do this too.

In your other window it was done long ago. This one just doesn’t know it yet.

What one wrong number costs

An illustration, not a customer story

You ask an ordinary AI to price an order. The supplier file says $84 per unit. The AI uses $48 – two digits swapped – adds your 20% margin and quotes $57.60 per unit. It says “done.” You send the quote and win the order.

500 units, sold at $57.60$28,800
500 units, bought at $84$42,000
Your loss, before a single hour of work−$13,200

The right price was $100.80. With Claims Need Receipts, every number needs a receipt: Claude has to show where it comes from – the file and the calculation – before it says “done.” A swapped digit shows up in the receipt, before you send the quote.

It’s not you. AI memory keeps what the AI itself decides is worth saving, not an exact record of what you agreed on. And nothing makes it prove what it claims.

You’re not alone. In a global study by KPMG and the University of Melbourne (48,000 people, 47 countries, 2025), 56% of people admitted making a mistake at work because of AI. And while 76% of small businesses in the US use AI, only 14% have it built into their core operations (Goldman Sachs 10,000 Small Businesses, 2026). Using AI isn’t enough. What’s missing is a system.

Sources: KPMG – Trust, attitudes and use of AI (2025) · Goldman Sachs (2026)

Most of this comes down to four habits AI doesn’t have on its own: proving, asking first, remembering, and not going in circles.

This version is for Claude (Cowork and Claude Code). Only use ChatGPT or another AI? The package is built so that, with small changes (for example a different file name instead of CLAUDE.md), other AI tools that read an AGENTS.md file could use it too. We haven’t tested that yet, so we don’t promise it.
Calculator

What it costs you today — and when the system pays for itself.

Your numbers, your math. Nothing is sent anywhere.

EXAMPLE — CHANGE TO YOUR OWN NUMBERS
You work 40 hours a week.
That’s about 4.7 h a week, 11.7% of your work. You’re truly productive for 35.3 of 40 hours.
Dead ends, redoing work, explaining things again from scratch. For comparison: one bad AI output costs about 2 hours on average (BetterUp + Stanford study, 1,150 people in the US, 2025).
ESTIMATE The author estimates 70% from his own work: without a system, AI wanders into dead ends; with it, it goes straight to the point and gets the work done faster. It hasn’t been measured on you — slide it down if you’d rather be more cautious.
Initial setup1 hour
Plan up to an hour: seven questions and running the installer. The system costs $199; we add this hour of your time to it as a cost, to keep the math honest.
What it costs you today
$8,640 a year18.0 hours a month
Productive hours per week
35.3 h → 38.6 htoday → with the system, at your estimate
The system pays for itself once it saves you
6.0 h of work5.0 h to cover the $199 price, plus the 1 hour you put into setup. Once, not every week. At your estimate you save 3.3 h a week, so you’ll reach those 6.0 h in about 9.1 workdays.
Even if it’s only half
(35%)		Your estimate
(70%)
Saved per week1.6 h3.3 h
Saved per year× 46 working weeks75.6 h151.2 h
Saved hours in money× your hourly rate $40$3,024$6,048
System priceone-time, tax included−$199−$199
Your hour of initial setup1 h × your hourly rate−$40−$40
Net gain, year 1$2,785$5,809
Pays for itself inuntil you’ve saved 6.0 h of work18.2 workdays9.1 workdays
Productive hours per week36.9 h38.6 h
This is a calculation from your numbers and an estimate, not a promise. Nobody can know in advance how much you’ll save — that’s why you have 14 days to get your money back.
With these numbers, the system won’t pay for itself in the first year, even at your estimate. Honestly: in that case, don’t buy it.
How we calculate itwork week = days × hours per day lost h/week = hours on a bad day × bad days a month × 12 ÷ weeks per year share lost = lost h/week ÷ work week productive today = work week − lost h/week costs you today = lost h/week × weeks per year × hourly rate break-even (h) = 199 / hourly rate + 1 h of setup saved h/week = lost h/week × your estimate 30–70% (half: × half your estimate) saved h/year = saved h/week × weeks per year in money = saved h/year × hourly rate system price = 199 your hour = 1 h × hourly rate net gain, year 1 = in money − 199 − your hour pays for itself (days) = (199 + your hour) / (saved h/week × hourly rate) × days per week productive w/ system = productive today + saved h/week

For comparison, from studies: at companies with revenue over $100 million, almost 40% of the time AI saves is lost again to reworking and double-checking results (Workday + Hanover Research, 3,200 people, 2025). People who use AI save about 2.2 hours a week on average (St. Louis Fed, 2024 data). These are not measurements of this system.

How it works

Four habits Claude doesn’t have on its own. With Claims Need Receipts, it does.

You get a ready-made file structure (rules, project state, change log and guard scripts) that the AI leans on. You don’t write any of it: your AI reads the guide in the package, asks you seven questions one at a time, shows you what it will change, and writes nothing until you say yes.

SCRIPT a guard script in the package checks it in your files · RULE a rule Claude follows on its own; a script can’t check it from files

01 · REMEMBERS

It remembers what you agreed on.

Show all 4
  • Picks up where you left off. It keeps decisions and where the work stands in files as it goes, not just at the end, so a new chat starts from them instead of from zero. SCRIPT
  • After a break, it re-reads first. The next morning, the rules tell Claude to read the latest state from your files again before it touches anything. RULE
  • Knows when a chat is getting too long. In Claude Code, a script measures how full the conversation is, and Claude offers a clean handover before it loses the thread. In Cowork this can’t be measured, and Claude says so. SCRIPT
  • Your rule files stay lean. Every limit lives in one settings file, and scripts check file sizes and split an oversized project state for you. SCRIPT
02 · PROVES

What it claims, it has to back up.

Show all 7
  • No number, date or “done” without a receipt. A claim about time, a number or the state of the work gets measured first, and the measurement leaves a line on disk. A script warns you when the receipt is missing. SCRIPT
  • The right date and time. Dates and times are measured fresh, never guessed, and always through your time zone, so a daylight-saving switch doesn’t fool it. SCRIPT
  • A refuted claim stops looking like a fact. When something it wrote turns out to be wrong, the rules tell Claude not to delete it but to mark it void wherever it appears, with the date, what applies instead and the proof. A script checks that every such mark is complete. SCRIPT
  • A task closes only with proof. Claude may close an open task on its own only with machine proof written next to it; without proof, only you close it. Scripts check that a closed task doesn’t still show up as open somewhere else. SCRIPT
  • You can see who changed what, and when. Every change gets a log entry: what, why, where, and how it was checked. A script checks that no required field is empty. SCRIPT
  • “Saved” on the screen isn’t proof. The rules list what counts as evidence and what doesn’t: a “saved” message, a badge or a filtered list is not; a fresh read from the source is. RULE
  • Mistakes get reported. When Claude makes a mistake while working, the rules say it has to tell you right away, fix it and write it down. RULE
03 · ASKS FIRST

It asks before any step it can’t undo.

GREENdoes it and tells you (reading, analysis, writing into the project)
YELLOWsays what it’s about to do, does it, reports back (moving files, editing several files)
REDwaits for your word (deleting, sending, publishing, payments)
Show all 6
  • Sending, paying, publishing or deleting waits for your word. Every action falls into one of three bands, decided by whether it can be undone and whether it goes out into the world. And one “yes” covers only the action it names – it’s never a blank check. RULE
  • Nothing gets deleted for good. The rules send everything Claude deletes to a trash folder inside your workspace, separate from your computer’s trash, and only you empty it. RULE
  • Passwords and keys stay out of the chat. The AI never enters, reads or prints a key – you put it in place yourself. A key that shows up in the chat is treated as leaked and replaced. RULE
  • It sticks to what you asked. If the task would reach into another project, the rules tell Claude to stop there, say where the work belongs, and wait for you. RULE
  • Your old rules don’t vanish. Before your instructions are rewritten, the rules make Claude back them up first, and a script checks that every version is labeled and logged. SCRIPT
  • It watches your usage limit and credits. On the first message about a usage limit or paid credits, the rules tell Claude to stop and ask you; it goes on with credits only with your consent and a cost estimate up front. It also tells you when a cheaper model would do. RULE

And before the installer changes anything, it makes a backup and can roll itself back.

04 · DOESN’T GO IN CIRCLES

It doesn’t send you down the same dead end twice.

Show all 3
  • It reads what already failed. At the start, the rules tell Claude to read what didn’t work last time; and when something fails, to say exactly what failed and offer a different way instead of trying the same thing again. RULE
  • “It can’t be done” only after it tried. Before it says something isn’t possible, it has to check its tools and record the attempt. A script flags a claim of impossibility with no attempt behind it. SCRIPT
  • Every answer ends with your next step. After a finished task, Claude ends with one clear line: what you need to do now – or that you don’t need to do anything. RULE
05 · THREADS AS A TEAM

Several conversations at once, working as one team.

  • Parallel chats without the chaos. Helper chats save their results to files, and the main chat checks each one with its own measurement before it tells you “done.” A script checks the register of running chats. SCRIPT
  • Two chats don’t overwrite each other’s work – usually. Right before writing, Claude re-reads a shared file and changes only its own section. A file mark is an agreement, not a lock: it doesn’t prevent two writes at the same time, it only helps catch them. RULE

Your AI names every new thread for you, for example “Website - 3 - 1 - prices” (the 3rd thread on the topic Website), and its helper threads “Website - 3 - 2 - …”, “Website - 3 - 3 - …”. Smaller tasks go to subagents: helpers inside the same thread, with no name and no new window.

To be fair: threads don’t see each other live; they hand work over through files. Subagents use your usage limit just like a new thread, and you can’t see into them while they run.

Where the scripts stop.

The package has 22 guard scripts and 10 helper tools. They check 92 of the 216 rules – the ones that leave a trace in files: times, numbers, log entries, hand-offs. The other 124 are about behavior in conversation; a script can’t verify those from files, so Claude holds them on its own. That includes most of the core: of its 14 rules, 4 have a script. Every mistake the system catches becomes a rule for your next work. We show which rules are checked in the 3D model below, instead of hiding it.

What it guards today · what you can let go

Start with receipts. Let go step by step.

You don’t have to trust the AI on day one. The system decides how much it may do on its own by one question: can the step be undone, and does it leave your computer?

RED · ALWAYS

Never on its own

Sending, publishing, deleting, paying, changing settings or accounts. Claude waits for your word every time. This is part of the core – no rule of yours can switch it off.

YELLOW · TODAY

Tells you first, then does it

Moving or renaming files, editing several files at once. It says what it’s about to do, does it and reports back – you see every step.

GREEN · TODAY

Does it and shows the receipt

Reading, analysis, drafts, writing into your project. Every claim about time, numbers and status still needs a measured receipt, and the scripts check what’s on disk.

Once the receipts keep matching for a kind of task, you can give it more room with a rule of your own – your rules win over ours. The red step stays red.

Why this exists

It’s great that you have AI. But your AI has no system.

I’m not a programmer. When I was twelve, I wrote a game in BASIC on a computer with a Zilog Z80: apples fell from the top, and I caught them in a cup at the bottom — and it was my first and last program. I trained as an electronics mechanic for computing and control technology, in other words computer electronics. I ran a consumer electronics repair shop (TVs, DVD players, VCRs, radios), and then spent 13 years putting together price quotes in Excel.

I worked at an installation company for 9 years and watched revenue grow while profit didn’t. It wasn’t the people — there was no system. Nobody knew how long a job really took or what it all cost. That led to my own project, ferovezakazky.sk — which makes me a good living today.

Then AI came along. For months it led me into dead ends — every piece of advice disconnected from the last. Tasks that AI couldn’t help me with, because it kept sending me around in circles, I had to hand over to a programmer. I waited weeks for one fix, and even that one I had to fix again with AI after it was delivered. It was a struggle without a system, let me tell you — it wandered through dead ends with me, but in the end we got it done together.

So I built myself a system with one single rule: if it isn’t written down, it isn’t true.

I spent almost half a year building the system, 8 to 10 hours a day. Everything in it gets written down, measured and checked. My AI still makes mistakes today. The difference is that they’re no longer invisible and, in my experience, they’re nowhere near as frequent — the system finds them and writes them down, each one becomes a lesson recorded in the rules, so the AI doesn’t keep repeating them the way it used to. The mistakes that cost me the most time were exactly the ones AI kept repeating. And on top of all that sit the check scripts. I’m 100% sure that with this system, AI makes far fewer mistakes.

Once the system was in place, things went differently: with it, Claude built license telemetry for me (ferovezakazky.sk) in a week. My programmer never even suggested it. And this package — the installer (Claims Need Receipts), 20 guards, the written guide and the legal texts — I put together with it in 8 days.

It works so well for me, and already saves me dozens of hours a month, that I decided to offer it to others. If I’d had it at the start, I would have saved myself months of dead ends.

— Branislav Koist, author

Rules without checks

Why Claude ignores your CLAUDE.md – and what checks it

Claude reads the rules in CLAUDE.md or in your project instructions. But unless you set up a check yourself, nothing verifies afterwards whether it actually followed them – and you only find out it skipped a rule from the mistake. You’re not alone: users report this publicly, right in the Claude Code developer repository.

Claims Need Receipts adds the missing step: guard scripts look at the files on your disk and print a warning where a receipt is missing – a number without a measurement, a change without a log entry, a hand-off without a trace. They check 92 of the 216 rules. The rest are rules about behavior in conversation, which the AI holds on its own. Which are which, we show below.

Public reports (GitHub, anthropics/claude-code): #17530 Claude Not Reading Claude.md · #19635 ignores CLAUDE.md rules repeatedly despite acknowledgment · #58642 does not enforce CLAUDE.md project instructions

The whole system, in 3D

216 rules. You can see which ones are checked.

Every dot is a real part of the package you get: rules, their families, the guard scripts that check them and the registers they read. The rules are general rules for working with AI, not tailored to my business. You add your own, and where they conflict with ours, yours win. Except the core: 14 rules in five areas no rule of yours can switch off. Nothing is deleted for good · red actions (sending, deleting, paying, changing settings) happen only on your word · the AI never handles passwords or keys · every claim needs a measured receipt · every change is logged. The grey rules have no guard.

216rules in the package you buy
45rule families
92checked by a script
124behavior rules — the AI holds them on its own
Loading the 3D model…

Coverage by rule family

Green = share of rules checked by a guard script. (The 3D view needs WebGL, which this browser doesn't offer.)

Why do 124 rules have no guard? A script can only check what leaves a trace in files: times, numbers, log entries, hand-offs. Rules about behavior in conversation (which language to speak, when to ask before deleting or sending, where the task ends) leave no trace on disk, so a script can’t verify them and the AI holds them on its own. For comparison: permission bands have a guard on 0 of 8 rules, language on 0 of 5, the change log on 7 of 8.
A computer checks it, not another AI. You know AI sometimes gets things wrong. That’s why an AI’s claims aren’t judged by another AI. Small programs (scripts) simply compare what the AI said with what is actually saved in the files on your disk: a date, a number, whether a task is really closed. It matches or it doesn’t. Nothing is guessed.

No virus, nothing hidden. The scripts are plain text files written in Python. You can open them in Notepad and have your own AI, or anyone, check them before you run them. None of them connects to the internet.

Drag to rotate · scroll or pinch to zoom · click a dot for details. The model shows structure and rule IDs; the wording of the rules is in the package.

What you’re buying

What you get, what you don’t, and what you need.

YESYou get

  • rules for Claude that load in every conversation
  • 22 guard scripts and 10 helper tools
  • templates: project state, handover, change log
  • rule files that stay lean: every limit lives in one settings file, and scripts check sizes and split an oversized project state for you
  • a written guide — why the rules look the way they do
  • an installer that deletes nothing, backs everything up and can roll itself back
  • updates released within 24 months
  • an optional mobile module (beta), installed switched off: check the state of your work and leave notes from the Claude app on your phone (your workspace synced with Google Drive)

NOYou don’t get

  • installation as a service, calls or consulting
  • an AI that never makes mistakes — nobody can build one yet; the system catches mistakes and writes them down
  • support for any AI other than Claude (not tested yet)
  • a team license — it’s for one person

NEEDYou need

  • Claude Cowork or Claude Code
  • a Windows or Mac computer (tested); Linux not tested yet
  • Python 3.10 or newer — if it’s missing, your AI walks you through installing it
  • up to an hour for the initial setup
  • the package files are in English — your AI talks with you in the language you choose
Price

One package. One payment. No monthly fees.

Claims Need Receipts
$199one-time

Tax included. No subscription. Updates released within 24 months included.

At your hourly rate from the calculator, it pays for itself after 6.0 h of saved work.
Buy once — $199

Payment is processed by Link as the merchant of record. Right after payment you get a personal download link on the thank-you page and the same link by e-mail; it works for 7 days and 3 downloads. Single-user license.
Requirements: Claude Cowork or Claude Code · Windows or macOS · Python 3.10 or newer (if you don’t have it, the guide walks you through installing it). Tested on Windows, macOS and Cowork (Cowork with Python 3.10.12); Linux not tested.

GUARANTEE ★ MONEY BACK ★ NO QUESTIONS ★ 14 DAYS

14 days. Money back. No questions.

Not a fit? Write within 14 days of your purchase, by e-mail or through the form, confirm that you’ll stop using the package, and we’ll refund the full amount, tax included, within 14 days of your request. You don’t have to give a reason. One refund per person.

Refund Policy

FAQ

Questions people ask

I’m not a programmer. Can I handle this?

Yes, that’s exactly who it’s built for. You don’t write or code anything: your AI reads the guide in the package, asks you seven short questions one at a time (one or two more if you already have rules of your own), shows you what it will change, and writes nothing until you say yes. Plan up to an hour. The guide that comes with the package explains everything you need.

Does it work with ChatGPT or other AI tools?

This version is for Claude (Cowork and Claude Code). The package is also built for tools that read an AGENTS.md file, but it hasn’t been tested with any other AI yet. Until we test it, we don’t promise it.

Will it stop every AI mistake?

No – and honestly, nobody can promise you that. Billions of dollars go into developing AI, and an AI that never makes mistakes doesn’t exist yet. This system is there to catch the AI’s mistakes and write them down so they’re far less likely to repeat: the AI knows about them in later work, and the most expensive ones are caught by a guard script (not the AI) when they repeat. Of the 216 rules, a script checks 92; the rest are rules about behavior in conversation, which a script can’t verify from files.

Claude says “done,” but it isn’t. How does this help?

The rules say that a claim about the state of the work, a time or a number has to be measured first, and the measurement leaves a trace on disk. Guard scripts (not the AI) then check that the receipt is there. Where it’s missing, they print a warning, so a “done” without a trace no longer looks like finished work. It won’t catch everything: a script checks 92 of the 216 rules.

Why does Claude ignore my CLAUDE.md or project instructions?

Claude reads your instructions, but unless you set up a check yourself, nothing verifies afterwards whether it actually followed them. This system adds the missing step: guard scripts look at the files on your disk and print a warning where a receipt is missing – for example a number without a measurement or a change without a log entry. They check 92 of the 216 rules; rules about behavior in conversation can’t be verified from files by a script, so the AI holds those on its own.

Can Claude remember decisions across chats?

Yes – in files, not in the AI’s own memory. In every conversation, the AI writes decisions and where the work stands into plain files on your computer. The rules file loads at the start of every new conversation, and the new conversation reads those files first – so you don’t start from zero and don’t explain everything again.

What if Anthropic changes Claude?

The system is built on today’s Claude (Cowork and Claude Code). Claude is developed by Anthropic, not by us. We can’t influence its features, limits or prices. When something changes, we’ll adapt the system in the updates included for 24 months, if it’s technically possible. We can’t guarantee, though, that Anthropic won’t limit or switch off a feature, and we’re not responsible for its decisions or for its AI’s answers. This does not affect your statutory rights.

Is it a subscription?

No. You pay $199 once, tax included. Updates we release within 24 months of your purchase are included.

How do I get the package after I pay?

Right after payment you land on a thank-you page with your personal download link, and we send the same link to the e-mail address you used for the purchase. The link works for 7 days and 3 downloads. Opening the page doesn’t use a download; only the Download button does. If the link runs out, write to brano@claimsneedreceipts.com from the address you used for the purchase, and we’ll open it for you again.

Will it overwrite my files?

No. If you already have your own rules (CLAUDE.md or AGENTS.md), they stay. The installer deletes nothing: it first shows you a dry run of what it will add, backs everything up, and adds only one marked block to each of those two files. Where your rules and the system differ, you choose during installation which one applies, and your choice wins – except for the core, 14 rules in five areas that always stay: nothing deleted for good, red actions only on your word, no passwords or keys handled by the AI, every claim with a measured receipt, every change logged. You can verify the installation and roll it back.

Is installation help included?

No. Claims Need Receipts is a digital product: files and a written guide. Installation as a service, a setup call or consulting are not included. If the package doesn’t arrive or the file is damaged, write to brano@claimsneedreceipts.com.

What if it’s not for me?

There’s a 14-day satisfaction guarantee. Ask within 14 days of your purchase – by e-mail from the address you used to buy, or through the online form – and confirm that you’ll stop using the package and delete your copies of it. We’ll refund the full amount you paid, tax included, within 14 days of your request. You don’t have to give a reason. The full terms are in the Refund Policy.

Is this made or approved by Anthropic?

No. Claude is a trademark of Anthropic. Claims Need Receipts is an independent product and is not affiliated with, endorsed or sponsored by Anthropic.

Will it burn through my Claude usage limit?

The rules work the other way: on the first message about a usage limit or paid credits, Claude stops and asks you; it continues on credits only with your consent and a cost estimate given in advance. It also tells you when a cheaper model would do the job. These are rules, not scripts – Claude holds them on its own.

Why does Claude need an architecture like this at all?

Because Claude doesn’t keep your whole history. Each chat has a limited working memory, long chats get summarized, and the built-in memory keeps only short notes – which can be incomplete or wrong. So the package moves what matters out of the chat and into plain files on your own computer: decisions, the current state of each project, what already failed, and every change. Your disk is memory you already own. A new chat reads those files first, so it works from what was actually decided – not from a guess. It doesn’t make Claude perfect, but it makes it less likely to get things wrong. (It can’t read your old chats – only what was written to the files.)

Why buy this – couldn’t I build it myself?

Yes, you can – and this site shows you most of the ideas. What you’re paying for isn’t the ideas, it’s the time behind them. This system was built over almost six months of near-daily work with Claude: work was logged on 140 of 174 days – since July, on all but two – and since August, when we started timing it, most days ran from morning to evening, with more than 1,700 recorded changes across 18 real projects. Most of the rules started as a real mistake made in that work; each one became a rule and, where possible, a script that catches it next time. Ask your own AI to rebuild it from this page and you’ll have a first draft in an hour – then months of finding, one by one, the mistakes we have already found and written down. Put a price on even a fraction of those hours and compare it with $199.

Learn more

Guides

Free articles in plain English that go deeper into the problems on this page.

All guides

Next time it says “done,” you’ll know where the receipt is.

Start by working out what this costs you today. Then decide.

Calculate Buy once — $199