A conversational AI bot is only as good as the instructions and knowledge it is given. This guide is about writing those well on the levelpartner.ai platform: the instruction block, the knowledge base, the guardrails and the test script. Get them right and the bot gives accurate, on-brand answers and books the right people. Get them wrong and no setting elsewhere will save it.
The four parts of a bot
Every bot on the platform has four parts you write:
- Instructions: who the bot is, who it talks to, its goal and its limits.
- Knowledge: the material it answers from.
- Actions: what it may do, such as book, tag or hand over.
- Tests: the conversations you run before and after any change.
Most bot problems trace back to one of these. Vague instructions produce vague answers. Thin knowledge produces guesses. Missing actions produce dead ends. No tests means you find out from a client.
Writing instructions that hold
Short beats long. Put the rules that must never break at the top, in plain sentences. Then the goal. Then style. A bot reads everything, but the clearest rules have the most effect.
A useful order is: identity in one line ("You answer enquiries for a 40-person consultancy in Leeds and Chicago"); must-never rules ("Never quote fees. Never give tax advice."); goal ("Find out the business problem and book a call"); style ("Two sentences per reply. British spelling. No lists."). Avoid personality descriptions; they add words without adding accuracy.
Before you start
- A one-page brief on what the bot is for, approved by whoever owns the channel.
- A list of topics the bot must refuse or hand over.
- Twenty real questions from your inbox, with ideal answers.
- Your knowledge material in clean, current form: no superseded documents.
- Access to AI Agents and the channels the bot will use.
How to set it up
- Go to AI Agents > Conversation AI and click + Create Bot.
- Select Prompt Based Bot.
- Choose Start from Scratch, or General Q&A for a support role.
- Continue into the bot editor.
- Configure Bot Settings with a name that says what it does.
- Configure Bot Training, adding your twenty short answers first.
- Configure Bot Goals using prompts, model settings and actions.
- Enter each test question in Test your Bot and check the reply.
Order the prompt the same way every time: identity, then the must-never rules, then the goal, then style. Add longer source material such as service pages only after the short entries are in, and switch on only the actions the goal needs. Your test script should hold ten conversations, two of which ought to trigger a handover and one that tries to make the bot break a rule. When a reply fails, fix it where it came from: an instruction for behaviour, a knowledge entry for facts. Then rerun the whole script, because a fix in one place can shift answers elsewhere.
Why short knowledge entries beat long documents
When a bot answers from a 20-page brochure, it has to find the relevant sentence among hundreds. When it answers from a two-line entry titled with the question, it finds the answer directly. Long documents are fine as background; the questions people actually ask deserve their own entries.
Picture a bot built on three long PDFs that answers about one in four test questions vaguely. Add 40 short question-and-answer entries, keep the PDFs as background and rerun the same test. If vague answers drop to around one in twenty, that would be five vague replies per hundred questions instead of twenty-five, without touching the instructions.
Time to maintain
Writing a solid first version takes a day. After that, the test script is the main ongoing cost: about 20 minutes to rerun after each change. Add a new test whenever a real conversation goes wrong, so the script grows with your experience.
People fix a bad answer by adding a new instruction each time. After a month the instructions are three pages of exceptions that contradict each other. Fix facts in the knowledge base and keep the instructions short.
When a bot is not the right tool
If your answers change daily, such as live stock or rates, a bot reading a static knowledge base will be out of date. Link to the live source or hand over to a person. If conversations are mostly emotional or high-stakes, keep a person in front. For a broader view of where bots fit, see conversational AI. For bots with a selling goal, the AI sales agent guide applies these principles to qualification, and AI customer support applies them to client service.
Troubleshooting
The bot breaks a must-never rule. The rule is buried or ambiguous. Move it to the top and rewrite it as a short, absolute sentence.
It answers correctly in tests but not live. Real visitors phrase questions differently. Add live phrasings to your test script and knowledge entries.
It contradicts itself across conversations. Two knowledge sources disagree. Find and remove the outdated one.
It hands over too often. The hand-off rule is too broad. Narrow it to specific topics and requests.
Frequently asked questions
How long should a conversational AI bot's instructions be?
Usually under 300 words. Anything longer tends to hide the important rules. Facts belong in the knowledge base, not the instructions.
Can one bot work on chat and phone?
The same knowledge can serve both, but phone needs its own instructions for pace and length. Most firms run a text bot and a voice bot from shared knowledge.
How often should we test?
After every change and at least monthly otherwise. Services, staff and prices change, and a monthly run catches drift.
Should the bot have a name and personality?
A name is fine. A detailed personality rarely helps and often hurts accuracy. Spend the effort on knowledge entries instead. The chat bot with AI guide covers greeting wording for websites.