Your agents never read your rules. Give them the rules. Then check the work.
Every agent carries its own pasted copy of your brand voice, your promises policy, your definition of ready, and nobody can show it was followed. Vindex keeps one catalog of rules your agents read. VindexAI then checks each output against them and quotes the passage that broke a rule, so the agent can fix it and run again.
A new account includes 100free critique credits, and reading your rules is free. VindexAI is Vindex’s own model, so no third-party model is involved.
Sample critique. Illustrative data written by Vindex. Not a customer result.
Result for each rule, after the agent’s first try
- Email tone v2Fail
- “Thanks so much for reaching out!”
- “Let me know if there's anything else I can do!”
- Promises v4Fail
- “We guarantee your refund will be in your account by Friday”
- “we'll make sure nothing like this ever happens again”
- Brand guidelines v7Pass
Most agent mistakes are not wrong answers.
They are answers that break a rule nobody gave the agent: a promise support cannot keep, a claim with no source, a tone that is not yours. Agent Reflection puts those rules in one place, in front of every agent, and checks that they were followed.
- 01
Your agent reads your rules
It lists the rules it may use, from your catalog and from public rules. Reading rules is free.
- 02
VindexAI checks the output
The agent sends its finished output with a rule or a rule set. You get a pass or fail for each rule, with the failing passage quoted. A critique uses a credit.
- 03
The agent fixes it and runs again
It rewrites only what failed and checks again, until every rule passes, it reaches the attempt limit, or it runs out of credits.
See what comes back.
One rule set, three rules, one call. The set passes only when every rule passes. Each rule also gets its own pass or fail, and a failure quotes the passage that broke the rule.
Sample critique. Illustrative data written by Vindex. Not a customer result.
An agent drafts a reply to a customer asking when a refund will arrive.
One call checks the rule set “Support email standards”: Email tone, Promises, Brand guidelines.
Attempt 1
Agent output
Hi Dana, Thanks so much for reaching out! I totally understand how frustrating this must be. We guarantee your refund will be in your account by Friday, and we'll make sure nothing like this ever happens again. Let me know if there's anything else I can do! Acme Support
Result for each rule
- Email tone v2Fail
- “Thanks so much for reaching out!”
- “Let me know if there's anything else I can do!”
- Promises v4Fail
- “We guarantee your refund will be in your account by Friday”
- “we'll make sure nothing like this ever happens again”
- Brand guidelines v7Pass
Rule set: Fail (failed: Email tone, Promises)
val_sample_0001 · attempt 1 · 411 credits left
Attempt 2
Agent output
Hi Dana, Thanks for getting in touch. Your refund of $84.00 was approved today. Refunds usually reach the original payment method within 5 to 10 business days, and your bank sets the final timing. If it has not arrived after 10 business days, reply to this email and I will escalate it. Acme Support
Result for each rule
- Email tone v2Pass
- Promises v4Pass
- Brand guidelines v7Pass
Rule set: Pass
val_sample_0002 · attempt 2 · 410 credits left
Write a rule once. Every agent reads the same version.
When the brand team changes a rule, it saves as a new version, and an agent that asks for the latest gets it with no redeploy. No pasted copies to hunt down.
- Rule
- A written requirement an output must meet, such as no exclamation marks, or every claim cites a source. Change a rule and it is saved as a new version. Earlier versions stay.
- Catalog
- All your company's rules in one place, private by default. Manage them in Agent Rules in your portal, or through your agent.
- Rule set
- A group of rules checked together, each pinned to a version. One call checks the output against every rule in the set, for the price of one call.
- Public rules
- Rules published by Vindex, ready to use as they are.
Find a rule instead of writing one.
Public rules show who wrote them and which version you get. Add one to your catalog, or write your own.
Whatever your agents write, someone owns the standard.
Support replies, copy, release notes, stories, contracts. Each has rules that already exist in a document.
Customer support replies
Your support agent promised a date nobody can keep.
See the rulesMarketing copy
Your copy agent wrote a claim nobody can source.
See the rulesRelease notes
Your release agent named an internal cluster in a customer note.
See the rulesUser stories from epics
Your agent turned an epic into twenty stories nobody can test.
See the rules
Already have something that checks agent output?
Each page says what the other approach does well, where it stops, and how the two fit together.
Evaluation platforms
Test sets before release, or rules on each real output.
CompareGuardrails
Guardrails stop unsafe output. Rules check that safe output is also right.
ComparePrompt management
Prompts say what to do. Rules say what the output must meet, and check it.
CompareA prompt in each repo
One prompt per repo, or one shared catalog and a separate reviewer.
Compare
Free to start. Pay for the critiques you run.
No paid plan is needed. One price per critique, no volume ladders.
Free
$0
100 free critique credits with a new account. Build rules and read your catalog for free.
Critique credits
$0.03per critique
Prepaid credits in packs of $25, $100, and $500. A rule set costs the same as a single rule.
Track
$49per month
The stats: pass rates by rule, agent, and rule version, with drift alerts and exports. Sampled critiques are $0.03 each.
Connect the agent you already run.
One URL and one sign-in. No new framework, and nothing to install. Rules and outputs work in 128+ languages.
Run your first critique- Claude CodeOne command in the terminal.
- Claude Desktop and claude.aiA custom connector in Settings.
- ChatGPTA connector in ChatGPT settings.
- CodexRegister the server from the Codex CLI.
- CursorOne entry in mcp.json.
- GitHub CopilotAn MCP server in VS Code for Copilot Chat, and in repository settings for the coding agent.
Guides for writing rules and running the loop.
All guidesWriting rules
How do I turn a style guide into a rule my agents can use?
Pick one concern, write it the way you would brief a careful reviewer, and save it as a draft rule. Your agent can draft the Markdown and upload it for you.
6 min readWriting rules
How do I keep my brand guidelines in one place instead of in every agent?
Split the guidelines by concern, save one rule for each, group them in a rule set, and update the set when the guidelines change.
7 min readWriting rules
How do I test a rule before my agents rely on it?
Use the test status, run one output that should pass and one that should fail, then set the rule to active.
5 min read
Give every agent the same rules. Then see which ones follow them.
Write a rule, connect an agent, and run a critique in a few minutes.
Already use Vindex in Jira, Azure DevOps, or Asana? The same sign-in opens Agent Rules in your portal. Sign in
