humble.fund

Best Practices for Setting Up an AI Assistant

Set up once, scale across your team. Build assistants that embed your best workflows, prompt engineering patterns, and collective knowledge.

Start SetupSee Example Slide Deck

Why It Matters

Designing a great AI assistant isn't just about answering questions — it's about structuring knowledge, instructions, and output so your team can solve problems faster, more consistently, and without needing to be prompt engineers.

By following this guide, you'll be able to:

  • Embed best practices from prompt engineering into reusable workflows
  • Scale knowledge and processes across your team
  • Continuously improve with usage tracking and iteration
  • Reduce repeated work while improving output quality

Visual Walkthrough

The Four Pillars of a Great Assistant Setup

1

Inputs (User Queries)

This is what your users will type into the assistant.

These are typically natural language questions or tasks they'd otherwise ask a colleague, such as:

  • •"What's our latest pitch deck for retail clients?"
  • •"Draft a follow-up email for someone who's gone quiet"
  • •"Summarize this customer call transcript"

You don't need to define these in advance — but understanding the kinds of requests you expect will help guide how you set up everything else.

2

Instructions (Short-Term Memory / Logic Layer)

This is where you define what the assistant should do with any input.

Instructions are used on every request. They define the assistant's role and the process it should follow. Be explicit, but not overloaded — too much here risks the model forgetting parts of it.

What goes in here:

  • The assistant's role
    e.g., "Act as a B2B sales specialist who writes concise, persuasive follow-up emails."
  • The processing logic
    e.g., "Take the input and turn it into a 2-paragraph summary, followed by 3 action points."

You can also create and name modular traits (e.g. "Tone: Friendly," "Output: Email")

Toggle traits on/off to A/B test different setups without changing the whole block.

3

Knowledge Base (Long-Term Memory)

Use this to give your assistant access to helpful background content. This information is searched only when relevant, and doesn't overload the assistant's short-term memory.

📂 What belongs here:

  • •Product documents
  • •Sales decks
  • •Pricing sheets
  • •Internal process guides
  • •Customer FAQs

💡 Example (Sales Enablement Assistant):

Knowledge base might include:

  • •Sales battlecards
  • •Case studies
  • •Feature comparison charts
  • •Email objection handling guide
🌀Auto-refresh can be toggled on for 24hr syncs
✅You can also manually refresh anytime. It's ready when you see the blue tick.
4

Output Format or Template

Templates help control the structure and quality of the assistant's response.

You can define:

  • •Desired tone
  • •Format (email, summary, LinkedIn post)
  • •Length limits
  • •Style examples

💬 #goodexample / #badexample pattern

#goodexample
Subject: Still thinking about us?
Body: Hi Alex — just checking in. Let me know if you'd like to pick things up again next week.
#badexample
Hey, are you ghosting me or what?

This trains the assistant to follow tone, structure, and language expectations.

5

Prompt Library (With Variables)

Prompt libraries let you create reusable forms of great prompts. You can use variables (via {{curly brackets}}) to turn them into fill-in-the-blank templates.

Example:

Draft an email for a client who has the following objection: {{What is the objection?}}

Users just answer the prompt field — the rest is handled by your predefined structure.

This improves user confidence and ensures consistency.

Troubleshooting & Tips

My assistant is giving weird or inconsistent answers
Check if any traits have conflicting instructions. Too many overlapping roles or tones can cause issues. Toggle off one at a time to test.
How do I update knowledge base content?
Open the folder → remove old files or upload new ones. Use manual refresh (blue tick confirms it's ready), or turn on auto-refresh for daily updates.
What if my assistant forgets instructions?
You're probably overloading short-term memory. Move background data to the knowledge base instead of packing too much into instructions.
Why modular traits?
They let you test and iterate in small pieces. Teams can tweak tone or format without breaking everything. Makes maintenance safer and easier.

Common Use Cases

Research Assistant

Market, legal, or competitor research with comprehensive data gathering and analysis.

Sales Enablement Bot

Access to decks, pricing, and customer objection handling in one place.

Internal Support AI

HR, IT, onboarding, and company policies made accessible through natural language.

Marketing Draft Helper

Create tweets, blogs, and emails from simple briefs with consistent brand voice.

RFP & Proposal Generator

Generate tailored answers from past documents and company knowledge base.

Your Setup Summary

  • 🔡
    Inputs:Understand common user queries
  • 🧠
    Instructions:Clear logic on what to do with inputs
  • 📚
    Knowledge Base:Searchable reference for deeper answers
  • 🧾
    Output Template:Define format, tone, and example answers
  • 🧩
    Prompt Library:Guide user input with reusable best-practice prompts

Build smart. Scale fast. And never write the same answer twice.