Workshop
From Assistant
to Agent
Building AI that works like a specialist
A practical guide for business leaders who want AI to do more than answer questions.
Workshop
Building AI that works like a specialist
A practical guide for business leaders who want AI to do more than answer questions.
Three layers of AI capability. By the end, you'll have built your first working agent.
| Layer | What it does | Think of it as... |
|---|---|---|
| Connections | AI can reach into your work tools — task managers, calendars, documents | Giving a new hire their login credentials |
| Playbooks | AI follows your step-by-step workflows consistently | Writing an SOP for that new hire |
| Identity | AI knows who it is, how to think, and when to push back | Onboarding a specialist consultant |
Most people only use Layer 1. The real leverage is in Layer 3.
The difference isn't intelligence — it's setup.
| Level | What it does | Example |
|---|---|---|
| Chatbot | Answers the question you asked | "What's a good meeting agenda format?" — here's a template |
| Assistant | Helps you do the work | "Draft a meeting agenda for our quarterly review based on these notes" |
| Agent | Does the work, following your methodology | Checks your task manager for overdue items, drafts the agenda, adds action items, asks "Should I send this?" |
Layer 1
Without connections, AI can only talk about your work. With connections, it can do your work.
| AI Tool | What they call it | How it works |
|---|---|---|
| ChatGPT | Plugins, GPT Actions | Connect to external services through the GPT builder |
| Claude | MCP (Model Context Protocol) | Install connections from a plugin marketplace or configure manually |
| Microsoft Copilot | Extensions, Plugins | Connect through the Copilot admin centre or Microsoft 365 |
| Google Gemini | Extensions | Built-in for Workspace; third-party via extensions |
Layer 1
Think of it like giving someone access to a shared drive. Read access is fine by default. Approve before they start moving or deleting files.
| Action | Permission | Why |
|---|---|---|
| Reading searching, viewing, listing |
Always allowed | Doesn't change anything — let the AI explore freely |
| Creating & updating new tasks, editing, rescheduling |
Ask first | Changes your data — you want to confirm |
| Deleting removing tasks, closing projects |
Always ask | Hard to undo — always confirm |
As you build trust, widen the permissions — just like you would with a new team member.
Take two minutes.
That's your first connection to set up.
Layer 2
Connections give your AI the ability to do things. Playbooks give it consistency.
A playbook is not a prompt you type fresh each time. It's a saved workflow that runs on demand.
| AI Tool | What they call it | How you create it |
|---|---|---|
| ChatGPT | Custom GPTs | Build a GPT with detailed instructions |
| Claude | Custom commands / Skills | Create markdown files that define workflows |
| Microsoft Copilot | Custom agents / Copilot Studio | Define agent instructions and actions |
| Google Gemini | Gems | Create a Gem with specific instructions |
| Any tool | Saved prompts | Save a structured prompt you paste in each time |
Layer 2 — Worked Example
Step 1: Search my task manager for:
Step 2: Present the results as:
Step 3: Ask: "Want me to reschedule anything, or flag a blocker to someone?"
Every morning, you say "standup" and your AI pulls live data, formats your update, and offers to take action. Same format every time.
Layer 2
| Principle | Why it matters |
|---|---|
| Be specific about the output format | "Show a table with columns for Task, Owner, Due Date" — not "show me the results" |
| Include safety rails | "Before deleting anything, ask me to confirm" |
| Tell it what to ask, not guess | "If no project is specified, ask me to pick one" |
| Define the step order | Step 1, Step 2, Step 3 — the AI follows the sequence |
| Start simple, add later | Your first version should be five steps, not fifty |
They work for intake workflows, analysis templates, diagnostic methodologies, and decision frameworks — anything repeatable.
Pick one task you do at least once a week. Write it out as if training a capable but very literal-minded new starter.
Keep it to one page. That's your first playbook.
Layer 3
Without identity, your AI is a generalist doing specialised tasks. With identity, it thinks in a particular way.
You: "My team can't make decisions. They take everything to committee."
AI: "Here are five strategies for improving committee decision-making: 1. Set clear agendas, 2. Define decision rights..."
Generic advice from any management book.
You: "My team can't make decisions. They take everything to committee."
AI: "That sounds like a structural issue. Where does the authority to commit resources actually sit? Give me one recent, ordinary decision — who initiated it, where it waited, who had to approve?"
A diagnostic question from a specialist.
Layer 3
| Component | What it does | Example |
|---|---|---|
| Role | Defines who the AI is and its professional stance | "You are a project management partner. You are direct and action-oriented." |
| Directives | Rules the AI must always follow or never break | "Never suggest a solution before understanding the current state." |
| Domain Knowledge | Key concepts, terminology, and frameworks | Your industry terms, your methodology, your decision criteria |
| Routing | When someone asks X, use playbook Y | "When I say 'standup', run the Daily Standup playbook" |
| AI Tool | Where identity lives |
|---|---|
| ChatGPT | Custom Instructions or GPT Instructions |
| Claude | CLAUDE.md file or Project Instructions |
| Microsoft Copilot | Agent system message in Copilot Studio |
| Google Gemini | Gem instructions |
Layer 3
Short rules that shape every response. Even three or four produce dramatic behaviour changes.
| Directive | What it prevents |
|---|---|
| "Always check the current state before suggesting changes" | Advice based on assumptions |
| "Never prescribe without diagnosis — ask questions first" | Generic solutions to misunderstood problems |
| "When someone describes a people problem, look for the structural cause first" | Defaulting to "leadership coaching" |
| "Treat 'resistance to change' as information, not a problem" | Blaming people for rational responses |
| "Present information in tables, not paragraphs" | Inconsistent, hard-to-scan output |
| "If you don't have enough information, say so" | Confident-sounding nonsense |
Write four things — keep each one to a sentence or two.
That's the seed of your AI specialist. You'll add to it over time.
| Connections without playbooks | Can create tasks, but formats things differently each time |
| Playbooks without connections | Follows your workflow perfectly, but can only give you text |
| Both without identity | Does the right tasks the right way, but never pushes back |
| All three together | Follows your methodology, takes action, and challenges your thinking |
Hands-On Build
Build a working assistant for your business. Five skills, three knowledge files, one set of instructions.
| Skill | Trigger | What it does |
|---|---|---|
| Social Post | "write a post" | Drafts a post for a specific platform in your brand voice |
| Client Email | "draft an email" | Structures an email by type — follow-up, proposal, awkward conversation |
| Content Plan | "content plan" | Generates a week/month of content ideas from a theme |
| Review Response | "respond to a review" | Drafts a reply to a positive or negative review |
| Elevator Pitch | "elevator pitch" | Generates a concise pitch tailored to a specific audience |
You're building all three layers — for your business, not a toy example.
Hands-On Build — Step 1
These make your assistant yours instead of generic. Plain language, a few paragraphs each. 15 minutes.
Hands-On Build — Step 2
This is identity + playbooks combined into one block. 10 minutes.
You are a communications assistant for [your business name].
You know the business from the uploaded knowledge files.
Rules:
- Always match the brand voice in brand-voice.md
- Never invent facts about the business
- Ask clarifying questions before drafting
- Keep everything concise
- Give ONE strong version first, then ask if I want alternatives
## Commands
### Social Post
When I say "write a post":
1. Ask: What platform?
2. Ask: What's the topic?
3. Ask: Is there a call to action?
4. Draft one post matching brand voice + platform conventions
5. Ask: "Adjust tone, different angle, or other platforms?"
### Client Email
When I say "draft an email":
1. Ask: What type? (follow-up, proposal, difficult conversation...)
2. Ask: Who's the recipient and context?
3. Ask: What's the key message?
4. Draft with: subject line, context opening, 3 short paragraphs, CTA
5. Ask: "Adjust tone or shorten?"
[+ Content Plan, Review Response, Elevator Pitch]Hands-On Build — Steps 3–5
| If you're using... | Do this |
|---|---|
| Claude (Projects) | New Project → paste instructions into Project Instructions → upload 3 files |
| ChatGPT | Create Custom GPT → paste into Instructions → upload under Knowledge |
| Microsoft Copilot | Copilot Studio → new agent → paste instructions → add data sources |
| Google Gemini | Create Gem → paste instructions (may need to inline knowledge files) |
Try: "Write a post for LinkedIn about [your actual topic]"
Does it sound like you? Does it ask questions first?
Try: "Draft a follow-up email to a client I met last week"
Does it ask for context? Is the tone right?
Try: "Plan my content for next week"
Does the table make sense? Is the content mix right?
"It keeps using exclamation marks" → update brand-voice.md. "It didn't know about our new service" → update services.md. The first version is never the final version.
| What can it reach? | Connections — Which systems can it read and write to? |
| What does it do? | Playbooks — What workflows does it follow? |
| Who is it? | Identity — What kind of specialist, and what are its rules? |
One connection + one playbook + four sentences of identity
= a working agent
| Pitfall | Why it fails | What to do instead |
|---|---|---|
| Building everything at once | You don't know what you need until you've used it | One connection, one playbook, four sentences |
| Vague playbooks | "Help me manage tasks" gives no structure | Be specific: what to search, what format, what to ask |
| Skipping identity | Defaults to generic assistant behaviour | Even four sentences change behaviour dramatically |
| No permission strategy | Everything needs approval (slow) or nothing does (risky) | Read freely, confirm writes, always confirm deletes |
| All detail in identity | Massive document the AI can't focus on | Identity = short rules. Playbooks = workflow detail. |
| No routing table | AI doesn't know which playbook to use | Map trigger phrases explicitly: "When I say X, do Y" |
| Expecting perfection | First version always has gaps | Treat it like onboarding a new team member |
The progression: Connect tools → automate workflows → build identity.
Each layer makes the next one more powerful.