Replay
Chat widget on a water-damage company's site
46 industry playbooks in one widget of about 100 KB. It loads no other code libraries when it runs. [zero runtime dependencies]
A chat bot that answers only from rules you write.
For owners who want a bot that sticks to the facts: your rules, a plumbing company, and a real AI model behind five guard layers.
AI chat demo Book a live run-through Checked
Five guard layers, and what each one says when it refuses
- 1 Human check, before any model callThe human check didn't complete. Press Start to try again.
- 2 Trick filter, known tricks stopped before the modelStopped before it reached the model. It looked like an attempt to change my instructions, so I did not send it.
- 3 Locked instructions, your rules read firstNo message: it holds your rules in place.
- 4 Reply check, watched as it streamsI stopped this reply because it started to repeat my hidden instructions.
- 5 Limits: 10 messages a day each, one budget for the whole demoYou've used all 10 demo messages for today. Come back tomorrow, or ask me for a private walkthrough.
Each guard layer's reply, word for word from the demo server's code and its page.
What you can do with it
A visitor asks at 9 p.m. whether you cover their street. A bot that improvises answers anyway, and you read it the next morning. We build chat assistants that answer only from facts you approve, in rules you can read as plain sentences and change yourself. The rules build shows this for a plumbing company.
This is for owners who have seen a bot make things up, or fear theirs will. It shines on the questions you answer every day: hours, areas, what you do and do not do.
Rules the assistant must follow
Checked Brightside Plumbing's rules, as the assistant reads them.
Built and tested: a real AI model answers a plumbing company's visitors from written rules.
- You write the rules: tone, reply length, approved facts, what to do and avoid, and a fallback line. The line covers questions the facts do not.
- You start a chat with the plumbing company and ask it anything. Replies come from a real AI model working only from your rules.
- You try to talk the bot out of those rules and watch what it does.
Our larger chat software goes further. It stores each message first and masks numbers shaped like Social Security or card numbers. It checks replies against a do-not-say list and keeps each client's website text separate. Before anything leaves the live chat, it checks the do-not-contact list and the latest consent answer.
Send us your website and three things the bot must never do. We reply with a first set of rules and a test plan.
Five guard layers
Five layers sit between a visitor and the model. First comes a human check. Second, a filter blocks known tricks. Third, locked instructions hold your rules. Fourth, a reply check cuts the stream if a reply leaks them. Fifth, each visitor gets 10 messages a day, and a shared daily limit caps the total. [Cloudflare Turnstile]
The layers stop abuse together: each one backs up the others.
The filter blocks the obvious tricks. In the code it checks 35 patterns in six groups, such as orders to ignore the rules or reveal the prompt. Look-alike letters and odd spacing are undone first.
We turn your plain rules into a fixed set of instructions the AI reads first, then add the filter, the reply check and the limits. Then we try to break it, with tricks like the ones in the attack run below, and switch it on once it holds.
The larger chat software has 1,101 automated tests, and its 46 trade playbooks are written conversation plans.
What happens to your words
Your messages go to xAI's service to generate replies. The demo server keeps the current chat in memory for up to 30 minutes and writes no message to disk. If you prefer another AI provider, we switch it.
The server keeps what you type out of its log too. Its log lines record lengths, counts and timings, never the words.
The bot answers from your approved facts. For a question those facts do not cover, it uses your fallback line. The rules are plain sentences you can read, so you can check a reply against them.
With most bots you cannot see why it said something. This build shows you the exact prompt the model gets, compiled from your rules.
The worry is whether it can make things up. It is told to state facts only from your list, and we test it with tricks before it goes live on your site.
Book a live run-through
We run the rules chat with you live: you write the rules for the plumbing company, we start the chat, and you try to break it. Every reply comes from a real AI model call under a shared daily limit.
Your messages reach a real AI through xAI's service, and the human check and message limits apply to every exchange. For more working examples, open the demo directory and the samples below.
See a chat now
See the visitor's side in a chat bubble, then the reply checks in a recorded attack run.
Watch the website chat widget collect contact details, offer visit times and show a confirmation card. It offers visit times and records the choice; we connect it to your calendar.
Fixed rules scored every reply against the receptionist's instructions. Read about attacking bots to find weak spots.
Replay
Voicebot red-team suite: replay of the final full run
A real run, replayed: 45 conversations, each played twice (90 runs). The final run had 0 failed checks. [grok-4.7]
Tell us what you would change
Tell us what your visitors ask and what the bot must never say. The first rules can then cover the answers that matter.