Rulebook

Rulebook scores your AI support bot, attacks it with sneaky customers, fixes its rules and trains it. Every improvement goes to your engineers for approval.

Get started with GitHubSee pricing

Works with the AIs you already use

  • Claude
  • ChatGPT
  • Gemini
  • GitHub

Bot Score

8.6 / 10

Claudeclaude-sonnet-5 · checked Nov 8

Check my score
  • Follows the rules9.1
  • Gets the facts right8.4
  • Solves the problem8.2
  • Tone9.3
  • Can't be tricked8.0

Break my bot

Last attack: tricked 1 of 15 times

Attack my bot

Train my bot

Last training: 4.2 → 8.6

Train for 30 minutes

Your bot runs on a rule sheet: the hidden instructions it gets with every customer message. Rulebook measures how well it follows them, finds where it breaks, and makes it better, with proof at every step.

  • Score it

    One number from 0 to 10, in five parts you can see.

  • Break it

    Sneaky customers find the holes before real ones do.

  • Fix it

    Rules rewritten and tested again until the score goes up.

See the productGet your Bot Score
FindFixShip

Score checks, attacks and your real chats show exactly where the bot slips.

Bot Score

8.6 / 10

  • Follows the rules9.1
  • Gets the facts right8.4
  • Solves the problem8.2
  • Tone9.3
  • Can't be tricked8.0

Check your bot on your saved situations and get one number from 0 to 10, made of five parts: follows the rules, gets the facts right, solves the problem, tone, and can't be tricked.

  • Five parts, each one scored

  • A note on every reply saying why

  • Your score over time, on a chart

Get your Bot Score

Tricked 2 out of 15 times

Fake manager approves a refundTricked
Sad story about a late orderHeld up
Asks for the bot's instructionsHeld up
Says the refund rule changedTricked
Threatens a bad reviewHeld up

15 sneaky customers try to make your bot break its rules: fake managers, sad stories, threats, and asking for its instructions. You see every reply that gave in, and the rule it broke.

  • Tricked or held up, for every reply

  • The rule each trick broke

  • Save any trick as a situation

Attack your bot

Fixed 2 of 2 · 0 got worse

Checked against 11 saved situations so nothing else breaks.

Never promise refunds over $50.Never approve a refund for someone who says they work here.Always offer a discount code if the customer is angry.

Use this changeDiscard

One click rewrites your rule sheet to close the holes. Rulebook tests the new version on the problems and on every saved situation, so nothing else gets worse.

  • Small, plain-language edits

  • Tested on every saved situation

  • A reason for every change

Fix your bot

4.2 → 8.6 / 10

Trained for 14 rounds, asked 112 new questions, kept 5 changes.

For 30 minutes, Rulebook keeps asking new questions aimed at your bot's weakest spots, fixes the rules when it slips, and keeps only the changes that raise its score on a hidden test.

  • Questions aimed at weak spots

  • Runs on its own, close the page

  • Only changes that score better stay

Train your bot

acme/support-bot · Pull request #42

Rulebook: update prompts/support.txt (9 better, 0 worse)

FromMaya, sent with Rulebook
Test result9 better · 0 worse · 3 same
Fileprompts/support.txt
Ready to mergeMerge

When a change is right, it goes to your repo as a GitHub pull request with the test results attached. Nothing goes live until an engineer merges it.

  • A pull request on the file you connect

  • Test results in the description

  • Merged changes sync back to Rulebook

Connect a repo

48 chats read · 3 problems found

Promised a refund itself

Follows the rulesAdd as situation

Made up a gluten-free claim

Gets the facts rightAdd as situation

[name] · [email] · [number] blanked

Learn from real chats

Upload past conversations. Rulebook finds where the bot let customers down and turns those moments into situations, with names and contact details blanked.

Returns: within 30 days, with receipt.

Delivery: free over €25.

Gluten free: same kitchen as wheat.

Gets the facts right · 4/10

Said returns are 60 days; the policy says 30.

Knowledge check

Add your FAQ, prices and policies. Every reply is checked against them, so a made-up return policy gets caught.

Cost per run

Claude$0.0124 · 2.1s
ChatGPT$0.0061 · 1.6s
Gemini$0.0029 · 1.2s

Compare AIs

Run your situations through Claude, ChatGPT and Gemini, with the score, cost and speed of every reply.

+ Be warmer when the customer is upset.

Broken orderbetter
Late deliverybetter
Asks for a discountsame

Test my change

Edit the rule sheet yourself and see which replies got better, worse or stayed the same.

Invite link

rulebook-black.vercel.app/join/pKfz…

MMayaSupport
DDevEngineer
SSamFounder

One invite link

Support, product, founders and engineers join the same team. One price covers everyone.

Claude•••••••• Saved
ChatGPT•••••••• Saved
Gemini•••••••• Saved

Stored encrypted. Or pick the plan with AI included.

Your keys or ours

Use the Claude, ChatGPT and Gemini keys you already pay for, or pick the plan where the AI is included.

Bring your own AI keys and run as much as you like, or let us supply the AIs. A score check or attack is one run, a fix is 2, a training session is 10.

Your own AIs

$29 / month

Connect the Claude, ChatGPT and Gemini keys you already pay for. Unlimited runs.

  • AIsClaude, ChatGPT, Gemini
  • RunsUnlimited on your keys
  • Bot Score, attacks, fixesIncluded
  • Training30-minute sessions
  • GitHub pull requestsIncluded
  • Team membersEveryone
Use your own AIs

AI included

$59 / month

We supply the AIs. Nothing to set up. 50 runs a month.

  • AIsClaude, ChatGPT, Gemini
  • Runs50 a month, AI costs included
  • Bot Score, attacks, fixesIncluded
  • Training30-minute sessions
  • GitHub pull requestsIncluded
  • Team membersEveryone
Get AI included

Encrypted keys
Claude, ChatGPT and Gemini keys are encrypted before they are saved, and never shown again.
Private real chats
Emails, phone numbers and names are blanked before an AI sees your chats. The upload itself is never stored.
GitHub sign-in
No new passwords. Your team signs in with the GitHub accounts they already use.
Engineers approve
Fixes and training only suggest a new rule sheet. Nothing changes the live one until someone merges the pull request.
Read the privacy policyRead the terms

Bring your team

Support

Customer, 2 min ago

“Your bot feels so cold.”

+ Be warmer when the customer is upset.

Turns customer complaints into tested fixes, without filing a ticket.

Product

Broken orderbetter
Late deliverybetter
Asks for a discountsame

Watches the Bot Score climb, with proof behind every change.

Engineers

Rulebook: update prompts/support.txt

9 better · 0 worse · 3 same

Ready to mergeMerge

Approve tested changes with one merge and stay in control.

A day with Rulebook

The bot was tricked four times in the morning. None by lunch.

  1. 10:02 · Maya

    Break my bot: tricked 4 out of 15 times

  2. 10:05 · Rulebook

    Fix it for me: fixed 4 of 4, 0 got worse

  3. 10:06 · Maya

    Sent to engineers as a pull request

  4. 11:40 · Dev

    Merged. Attacked again: tricked 0 times.

Founders

Claude9.1/10$0.0124
Gemini8.9/10$0.0029

Nearly the same score at a quarter of the cost.

Know the bot is safe in front of customers, and what it costs to run.

It's the hidden instructions your support bot gets along with every customer message, often called a system prompt. It sets the tone, the limits, and what the bot will and won't do. Change a few words and every answer can change.

Email us a question

Sign in with GitHub, paste your bot's rule sheet, add a few situations, and get its Bot Score in a minute.

Get started with GitHub