A rule is not enforcement

What it is

Most teams have a list of corrections they give their AI assistant over and over. Never promise guaranteed results. Always number the choices. Stop using the old campaign name. The rule is written in the instructions, and the assistant still breaks it.

Reply gate turns that list into a check. Each rule has a pattern to look for, where to look, and the reason the assistant will see. A reply that breaks a rule never reaches you. It goes back with the reason, and the assistant fixes it. The zip ships with eight invented example rules to copy and edit.

Who it is for

Marketing leads and operators who use an AI assistant every day and keep repeating the same corrections. If your style guide lives in a document the assistant can skip, this makes the parts that matter impossible to skip.

How to run it

You need Python 3.8 or newer. Nothing else to install.

1. Download the zip and unpack it.
2. Copy rules.example.json to rules.json and edit the rules.
3. Try it on a saved reply: python reply_gate.py --check reply.txt. It lists every rule the reply breaks. A result of 1 means at least one rule was broken.
4. To check every reply inside Claude Code, add it as a Stop hook in your Claude Code settings. The README has the exact lines to paste.

Before you switch a rule on, run it over a batch of past replies. If it fires on good replies, tighten the pattern. A rule that fires too often teaches people to ignore it.

Licence

Reply gate is released by Unicorn Flames LLC under the MIT licence. You can use it, copy it, change it and share it, including at work. The licence file is in the zip.

Frequently asked questions

Is it free?

Yes. Reply gate is free to download, use, copy and change under the MIT licence.

Does my team need to know how to code?

No. Someone needs to edit one rules file and run one command, and the README walks through both. If your team works in Claude Code, Claude can set it up for you.

Does it send my replies anywhere?

No. It runs on your own computer and reads only the reply it is checking. Nothing is uploaded.

Can it get my assistant stuck in a loop?

No. After two blocks in a row in the same session it lets the reply through and logs it. If the rules file is broken, replies go through untouched.

Is it a security tool?

No. It checks habits, not safety. Use it to stop repeat mistakes in tone and wording, not to protect private data.

Want your team's AI to follow the rules you set?

We teach marketing and communications teams to use Claude well, no developer required. Checks like this are one of the habits we set up with teams.

See AI training for marketing teams . Book a 30-minute call