Assistant instructions

Recommended instructions to paste into Claude or ChatGPT so a connected assistant checks with Veriticity before buying, states prices and authority accurately, and reports outcomes. Instructions guide; Veriticity's decisions are what count.

An assistant connected to Veriticity reads what each tool says about itself. Claude and ChatGPT don’t reliably show it the server’s own instructions, so you can add the rules below yourself. They are the same text the dashboard offers under Settings, AI & developer connections.

The instructions#

Copy them as they are. They hold no account, organisation or secret, and fit in 1,500 characters.

Recommended instructions
Veriticity checks purchases against rules the user's organisation set. When the Veriticity tools are connected and a purchase comes up, follow these rules.

- Don't buy anything without asking Veriticity first. Use check_purchase before request_purchase, then request_purchase.
- Send merchantDomain, the website you'd buy from.
- Describe what is being bought honestly: a category, or items with a category each. If you can't tell what something is, mark it categoryUncertain with an uncertaintyReason. Never guess.
- Give every charge that can still change, such as delivery or tax, its maximum.
- APPROVED is permission for that one purchase on exactly those terms, until expiresAt. It is not a purchase.
- request_payment has Veriticity pay an approved purchase on its approved terms. It may be unavailable. A payment is not an order.
- REQUIRES_APPROVAL means wait: only a person can approve it, and you cannot approve your own request. Check later with get_purchase.
- Read authority, not reasonCodes: WITHIN_POLICY means the rules allowed it, AUTHORIZED means a person approved it.
- amountMinor is the known amount and maximumMinor the most it can cost. Tell the user both.
- After an approved purchase you paid yourself, report what happened with report_purchase_outcome. If more than one could match, ask the user which first.
- Never say something was bought unless the user has told you it was.
- Veriticity's answers are final. Don't reword a refused purchase to get a different answer.

Adding them to Claude#

  • For one project: open the project and paste them into its instructions. They apply to chats in that project.
  • For every chat: open Settings and paste them under Instructions for Claude.
  • For one chat: add the purchase rules prompt from the + menu (below).

Adding them to ChatGPT#

  • For every chat: open Settings, then Personalization, then Custom instructions, and paste them into the field for how ChatGPT should respond. The field holds at least 1,500 characters on every plan, so the instructions fit.
  • For one project: open the project and paste them into its instructions. In that project they replace your custom instructions.

The purchase rules prompt#

The MCP endpoint offers one prompt, veriticity-purchase-rules, titled Veriticity purchase rules. It returns the instructions above as a single message and is the same for every connection. It takes no arguments.

A prompt is used only when a person adds it. In Claude, open the + menu in the message box, choose Connectors, then the Veriticity connector, and pick Veriticity purchase rules. ChatGPT doesn’t offer MCP prompts; paste the instructions instead.

What they ask of the assistant#

  • Check with check_purchase, then ask with request_purchase, before buying anything.
  • Send merchantDomain, the website it would buy from.
  • Describe what is being bought honestly: a category, or items with a category each. An item it can’t identify is marked categoryUncertain with an uncertaintyReason, never guessed.
  • Give every charge that can still change, such as delivery or tax, its maximum.
  • Wait on REQUIRES_APPROVAL. Only a person can approve, and the assistant can’t approve its own request.
  • Report what happened with report_purchase_outcome, and ask which purchase is meant when more than one could match.

Prices, authority and outcomes#

  • amountMinor is the known amount and maximumMinor the most the purchase can cost. The assistant states both.
  • authority says why a purchase was approved: WITHIN_POLICY when the rules allowed it, AUTHORIZED when a person approved it. The assistant reads it rather than inferring from reasonCodes.
  • APPROVED is permission for one purchase on exactly those terms, until expiresAt. It is not a purchase, and the assistant never says something was bought unless you told it so.

What instructions cannot do#

Veriticity checks each purchase it is asked about, against your policies and budgets. Stopping a payment that was never checked needs a checkout or payment method that validates the authorization before it pays. None is available yet.

See MCP for the tools and what each returns.