Chapter 02

Train your AI stand-in like a new teammate.

A useful stand-in needs more than facts. It needs judgment: what your company does, how you talk, what the page is for, when to ask for contact details, and when a real person should join.

Get your StandRead the chapter

Stand Guidebook

Chapter 02 of 09

Field guide

What to learn in this chapter

Write instructions the way you would brief a capable new hire on their first day. Start with behavior and boundaries, then add deeper sources only when visitors actually need them.

Use this chapter when you are writing the first prompt, splitting knowledge between team, site, and agent context, or deciding whether Business knowledge bases are worth adding.

Put shared behavior in the team prompt.

Use site prompts only when context changes by page or domain.

Give each agent one clear job.

Test the full stack before turning the agent on.

Prompt design

01

Put behavior in prompts and reference material in sources.

Prompts should teach judgment: audience, tone, claims to avoid, handoff rules, contact-capture timing, and what a successful conversation looks like.

Knowledge bases should carry longer reference material: docs, policies, product catalogs, implementation guides, comparison notes, and content that changes too often to paste into every prompt.

Stand agent identity screen showing internal name, display name, greeting, specialization, team prompt, and training prompt.
Figure 2-1. The Identity tab is the brief: visitor-facing name, greeting, site specialization, inherited team prompt, and the agent-specific training prompt.

Example

The pricing agent

Situation
Visitors ask which plan fits and whether a competitor would be cheaper.
Move
Put pricing principles and escalation rules in the agent prompt. Put the full plan matrix in a source or page content.
Insight
The agent should not memorize every exception. It should know how to guide a decision and when to invite a human.

Example

The docs agent

Situation
Visitors ask specific technical questions, but the source of truth is already in docs.
Move
Use the site prompt to say this is a developer context. Attach docs as a knowledge base on Business when the prompt starts becoming a mini manual.
Insight
A docs agent should answer the question and learn why the visitor needed it.

Prompt pattern

Weak prompt, better prompt

Weak
Answer questions about our product and collect leads.
Better
Help evaluators decide whether Stand fits their situation. Answer from the attached sources, ask one clarifying question when the use case is vague, and offer a human when the visitor has a live project or migration concern.
Why it works
The better version defines audience, behavior, uncertainty, and handoff policy. It gives the stand-in judgment, not just a task.

Knowledge bases

02

Use knowledge bases for source material the agent should retrieve.

Knowledge bases are Business plan stores for longer factual material. They are shared across your organization, then attached to specific agents when those agents need retrieval during a visitor conversation.

Use them for support manuals, docs, policies, implementation notes, product catalogs, or website sections that are too long or too changeable for a prompt. Do not use them to bury behavior instructions. The agent still needs a clear prompt that explains when and how to use the source.

App link

Open Agents in Stand.

Use the Agents page to create knowledge bases, upload documents, crawl websites, and attach selected knowledge bases to an agent.

Open Agents
Stand Knowledge bases screen with shared organization knowledge bases, document upload, website crawl, source status, and extracted text preview.
Figure 2-2. Knowledge bases are organization-level sources. Upload documents, crawl a URL path, inspect processed sources, then attach the base to the agents that should retrieve from it.

Before attaching a knowledge base

  • The material is factual source content, not behavior guidance.
  • You have the right to use the uploaded or crawled material.
  • The name is source-specific, such as Product Support Manual.
  • The source shows completed status.
  • The extracted text looks clean enough for retrieval.
  • The agent prompt tells the stand-in when to rely on the knowledge base and when to ask a human.
UI signalStorage 676.2 KB / 50.0 MB
What it meansBusiness storage is counted from extracted text, not just original file size.
Reader moveWatch storage before crawling large docs sites.
UI signalUpload document
What it meansAdd PDF, Word, Markdown, or text files such as manuals and policies.
Reader moveUpload source-of-truth documents instead of pasting them into the prompt.
UI signalCrawl website
What it meansAdd a URL and Depth. Depth 0 crawls only this page; higher depths follow links under the same URL path.
Reader moveStart shallow, inspect the source, then increase depth only when needed.
UI signalcompleted
What it meansThe source has finished processing and can be retrieved when attached to an agent.
Reader moveWait for completion before judging answer quality.
UI signal0 agents / 4 agents
What it meansHow many agents currently use that knowledge base.
Reader moveAttach the KB to the agents that need it; leave it unattached if it is only being prepared.
UI signalchunks / vectors
What it meansThe extracted text has been split and indexed for retrieval.
Reader moveInspect extracted text when answers seem off.

Layering

03

Avoid repeating yourself in every agent.

If every agent needs the same company description, tone, privacy boundary, or handoff rule, it belongs in the team prompt. If the instruction only applies to one domain or audience, put it in the site prompt. If it defines a specific stand-in job, put it in the agent prompt.

This keeps the system easy to change. When your positioning changes, update the team prompt once instead of hunting through every agent.

LayerTeam
Best forCompany voice, broad rules, universal handoff policy
Avoid putting herePage-specific pricing or one-off campaign notes
LayerSite
Best forDomain, region, product area, page audience
Avoid putting hereGeneral company story repeated from the team prompt
LayerAgent
Best forRole, greeting, specialty, success criteria
Avoid putting hereLarge documents or frequently changing source material
LayerKnowledge base
Best forLong docs, PDFs, website crawls, policies
Avoid putting hereTone, behavior, or what the agent should do next

Testing

04

Test for judgment, not trivia.

Try it out with questions that reveal whether the stand-in understands the job. Ask a pricing edge case, a skeptical comparison, a vague "is this for me?" question, and a question it should not answer confidently.

A good test chat ends with a better prompt. Tighten what the agent should ask, what it should avoid, and what should trigger a human offer.

Stand Try it out tab showing a test chat with an AI agent answering a technical product question.
Figure 2-3. Try it out is where you test the agent before putting it in front of traffic. Use it for judgment questions, not just factual recall.

A useful test set

  • One normal buyer question.
  • One vague question where the agent should ask a clarifying follow-up.
  • One competitor or alternative question.
  • One question outside your current promise.
  • One moment where the agent should ask for contact details or offer a human.

Questions

Common reader notes

Do I need a knowledge base to start?

No. Many teams should start with team, site, and agent prompts. Add knowledge bases when visitors need longer source material.

What formats do knowledge bases support?

Business knowledge bases can use PDFs, Word documents, Markdown, text files, and crawled website URLs.

Continue the guide

Build Stand as a learning loop, one chapter at a time.

Previous chapter 01

Start Website Conversations

Turn silent website traffic into useful chats.

Next chapter 03

Spot the Conversations That Matter

Define the signals your team should hear about.

Try the guide on one real page.

Get your Stand - free forever