A practical guide for people who build with AI
Your AI built it.Now make it safe to launch.
Before the First Customer names the failures that AI-built products keep repeating, gives you the exact instruction to ask your assistant for each fix, and shows you how to check the work.
Buy the book
£29.99. Tax is worked out at checkout from where you are. Payment and tax are handled by Paddle, who sell the book to you. You can ask for a refund within 30 days. See the refund policy.
Concept Informatics Ltd, VAT number GB 746 4847 93.
Not ready to buy?
Leave your name and email address and we will write to you about the book. We will not use it for anything else.
An assistant builds what you ask for
It rarely builds what you did not ask for, and almost everything that keeps a product safe falls into that second group. The demo looks identical either way. The difference only shows when someone behaves badly, when something fails at a bad moment, or when the product grows past the size of the test data.
A number in the address bar
A customer changes 4821 to 4822 and reads someone else's invoice.
A button clicked twice
One slow export, two clicks, two exports, and a customer with duplicates.
A link that goes somewhere else
An emailed sign-in link sends people on to a site you do not own.
These are invented examples of common failures. None describes a real company or event.
What is inside
Fourteen chapters, each on one failure that keeps happening, plus two appendices you can use without reading the rest.
Secrets
Keys that outlive their welcome, and how to replace them without an outage.
Who is asking
Making sure each customer sees only their own data.
Signing in
Sessions, tokens and emailed links that cannot be turned against you.
The admin door
Why a hidden link protects nothing, and what does.
What your errors say
Telling users little and your logs everything.
The database
Pages that fall over at scale, changes that lock, backups that never restore.
Slow work
Background jobs, retries and asking twice doing the work once.
Public doors and floods
Forms that robots use, and the three layers that stop them.
Money and webhooks
Forged payment notices, duplicates and the day the provider is down.
One customer's yes
Saying yes to one customer without charging all the others.
What you run
Pinning, inventories and what to do when a project changes hands.
A model near your data
Prompt injection, narrow actions and logging every call.
Privacy, paperwork and data location
Keeping what you say true, including where your data lives and whose law applies.
Launch and after
Scanning yourself first, support playbooks and a monthly rhythm.
Standing instructions
A file to paste into your assistant so it applies every time.
Pre-launch checklist
One page, with a place to record the evidence for each line.
How each chapter works
- A story. A short, invented example of how the failure unfolds.
- Why it happens. The reason an assistant, and the person guiding it, ends up there.
- The fix. The practice that prevents it, in plain steps, with the reason behind each one.
- Ask your assistant. An instruction you can paste straight in.
- Check it yourself. A few minutes of work to see whether the fix is real.
- The checklist. A few lines to keep.
A taste of the instructions
Every chapter has instructions written to be specific, because a vague request gets a vague result. This one is from the chapter on access.
Audit every route that takes an id (in the address, query or body). For each one, show me the line that checks the record belongs to the signed-in user or their tenant. List the routes where there is no such check. Do not fix anything yet.
And the check you run yourself afterwards: sign in as a customer in one organisation, copy the address of a record, then paste it while signed in as a customer in another. You should see "not found".
Who it is for
- Founders, developers and small teams shipping products built with AI assistants.
- Agencies and freelancers who deliver AI-assisted work to clients and need a standard to point at.
- People buying or inheriting an AI-built codebase who need a way to judge it.
You do not need a security background. You do need to be willing to read a few lines of code or a settings page when the book asks you to.
Who it is not for
- Security specialists. You will know most of it.
- Anyone looking for legal advice or a penetration test. The book is neither.
- Anyone hoping to be told a product is secure. No book can do that.
Questions
Do I need to be a developer?
No, but you need to be comfortable asking an assistant to show you code and settings, and reading what it shows. The book tells you what to ask and what to look for.
Does it work with any AI assistant?
The instructions are written in plain language and you paste them in. They work with assistants that can read your code and follow written instructions. Where an assistant can read a project instructions file, Appendix A gives you one to use.
Will it make my product secure?
No. It makes the most common and costly mistakes much less likely, and gives you a way to check. Security and reliability are large subjects, and the book is not an audit.
Is it legal advice?
No. Where a chapter touches law, contracts, payments or privacy, it describes the questions to put to a qualified person. The privacy chapter is written with the UK and EU in mind.
How do I buy it, and what does it cost?
Press the Buy the book button. It costs £29.99, with tax worked out at checkout from where you are. Paddle takes the payment and sells the book to you. If it is not useful to you, ask for a refund within 30 days. The refund policy has the details.
Can my team use it?
Yes, there is a team licence. Register your interest, or email info@conceptinformatics.com, if you need a different number of people.