Konradodocs

Working with Konrado

What your AI app is told when it connects, where each kind of fact goes, and three ready-made prompts.

When your AI app connects to Konrado, Konrado sends it working rules along with the tools. The rules say how to read and change your agents' setup without breaking it. This page tells you what they say, so you can check what your app proposes.

Konrado sends the same rules to every app. What an app does with them is up to the app. What an app can never go past is the access you approved when you connected.

What the rules say

  • One connection works with one organization: the one you picked when you signed in.
  • The app reads the setup before it changes it: each agent's instructions, guardrails, communication style, example replies and auto replies.
  • It changes only what you asked for. Some tools change one part, such as one guardrail or one example reply. Others replace a whole text or list.
  • Every change lands in Recent activity on the MCP Connection page, with the text it replaced.
  • To see how your agents do, the app reads the Insights numbers, the replies your team changed before sending, and the questions your Knowledge Base does not answer yet.
  • When no tool can do what you ask, the app says so and offers to report it to the Konrado team.

The rules also say where each fact goes, the limits, what never goes into the setup, and when to switch to Live. The sections below cover each.

Where each fact goes

Each kind of fact has one home. If an entry already covers a fact, the rules tell the app to change that entry instead of adding a second one.

What the agent should knowWhere it goesWhere you see itExample
What is true on every ticket: who you are, what the agent handles, when it hands over to your team, the links it can trustInstructionsAgentsInstructions"We host websites and domains. Refunds go to the billing team."
Something the agent must never say or doA guardrail, one sentenceGuardrails, on the same page"We do not offer free trials" becomes the guardrail "Never offer a free trial"
How a reply reads: greeting, length, tone. Never what it saysCommunication style and Example repliesAgentsHelpdeskAI instructions"Open with the customer's first name. Keep it short."
The steps for one situationA procedure. Its trigger starts with "When the customer"AgentsProcedures"When the customer asks to move a domain to us": ask for the transfer code, then check the domain is unlocked
One question customers ask, answered in one or two sentencesA Q&A entryWorkspace SettingsKnowledge Base"Do you move my website for free?" "Yes, one website on every plan."
A question only your team can answerA Q&A entry with no answer, and why Konrado asksThe same list, marked Waiting for your answer"Do you back up mailboxes, and for how long?"
A reference list: plans, prices, fees, policiesA Text sourceWorkspace SettingsKnowledge BaseYour price list for every plan
What Konrado learned about one customer or from your teamMemory. An app can correct a memory or remove it, never add oneAgentsMemoryA memory says refunds take 14 days. They now take 7: ask the app to correct it

A rule that starts with a situation is a procedure, not part of the instructions. "When a customer asks for a refund, check the invoice first" goes in a procedure.

A question waiting for your answer is not used in replies. Open it from the list, read Why we are asking, write the answer and click Save answer. Konrado starts using it as soon as you save it.

Limits

The rules give the app the same limits the console has:

WhatLimit
Instructions30,000 characters
Communication style rules plus Example replies30,000 characters per help desk
Guardrails20 on each channel, the agent's and the channel's own together
A procedure's trigger500 characters
A Q&A answer2,486 characters

No names or secrets in the setup

The rules tell the app never to write these into the setup:

  • the name, email address or domain of one of your customers
  • the name of someone on your team
  • a key, a password or any other secret

The setup reaches every reply. It holds what is true for all your customers, not details about one person. The Review last week prompt says the same for what it proposes: it tells the app never to copy a customer's name, email, domain, license key or password from a ticket into a change.

Sending replies on its own

On Live, Auto replies send replies to your customers when Konrado is confident enough, and nobody on your team checks them first. So the rules tell the app to switch a help desk to Live only when you ask for exactly that.

To see what auto replies would do before a customer does, ask for Shadow. Konrado decides on every ticket and sends nothing. Try auto replies in Shadow mode.

Three ready-made prompts

A prompt is a ready-made request Konrado gives your app. The app runs it like a request you typed.

PromptWhat it doesWhat it changes
Review last weekReads the drafts your team changed a lot in the last 7 days, and which procedures ran on those tickets. It finds why each draft missed: a fact the agent did not have, a wrong fact in the setup (such as a Text source that gives the wrong fix), a wrong or missing rule, a procedure that did not match, the style, or work only your team can do. It groups the misses by cause, the biggest first, and proposes the change for each cause, one edit or several, with the exact text and where each edit goesNothing until you pick. Then only the changes you picked
Answer waiting questionsGoes through the questions waiting for your team's answer, oldest first. For each one it shows the question and why Konrado asks, then waits for your answer. You can skip oneSaves each answer you give, in one or two sentences, in the language of the question. It shows you the text before it saves
Agent healthThe last 30 days in five lines: tickets, how many replies Konrado sent on its own, draft quality and its trend, the weak drafts of the last 7 days, the biggest gaps in your Knowledge Base, and how many questions wait for your answer. Then the three changes worth making firstNothing

Run a prompt

In Claude Code, type /. Claude Code lists a connected app's prompts as /<server name>:<prompt name>. The Claude Code command on the MCP Connection page adds Konrado as konrado, so the three prompts are:

/konrado:review-last-week
/konrado:answer-waiting-questions
/konrado:agent-health

If your app lists prompts from its connected apps, look for these three there.

Any app can do the same work from a plain request. Ask in your own words, for example:

  • "Review how our Konrado agent did last week and propose changes. Change nothing until I pick."
  • "Show me the questions Konrado is waiting for my team to answer, one at a time, and save my answers."
  • "Give me a health check of our Konrado agents for the last 30 days, in five lines. Change nothing."

Last updated on

On this page