Project Nebula · Session 2 · Building Self-Learning Agents with Claude Code

PM Workshop · Module 3 · Session 2

Project Nebula

Building self-learning agents with Claude Code

Continuation from Session 1 — context, memory, MCPs, skills, and custom tools.

Feb 2026 · Session 2

Before we start

Pre-requisites

Setup checklist

Claude Code installed, with Anthropic Console login completed.

One data-warehouse connection — Kusto / ADE in this example, but the workflow applies to Snowflake / BigQuery / Databricks too.

2–3 sample queries from that warehouse — copy them from one of your dashboards.

An email client for the email-skill exercise — Outlook (classic) for COM automation, or Gmail API / Apple Mail.

The agent may install libraries (Python, etc.) in your environment as needed.

⏱

As always — lots of patience

First-time installs and MCP wiring take a few minutes each.

Finding connection details

① "More actions" next to your cluster

② "Edit connection" → copy details

Your warehouse UI will look similar.

Claude Code · Data Warehouse · Email Client · Patience

Reminder

Graduation project

Identify tasks in your work that can be supercharged by introducing agent workflows.

Build an agentic system that can be leveraged by you or your team to improve productivity.

Claude Code

Copilot Studio

Or any other platform

Submit

Drop your ideas in the Module 3 — Agent graduation project board.

Continuation from Session 1 · Agent Series

Recap

We learned…

Definition

An agent is a system that pursues a goal by repeatedly:

01

Reasoning / Planning

→ Deciding what to do next

02

Acting on Environment (Tools)

→ Executing actions via tools

03

Updating Itself (Context/Memory)

→ Learning from results

What we built

An agent on Copilot Studio that had access to tools and reasoned over user requests to plan and use those tools to achieve the goal.

What was missing

Our agent could not learn from interactions and get better over time.

That's what Session 2 fixes.

What we'll cover today

Session 2 · Learning objectives

Concepts

Context & Memory

Understanding agent context, context window, and context engineering.

How agents can save context as memory.

Walkthrough

Introduction to Claude Code

Installation and walkthrough of Claude Code.

Best practices.

Self-guided exercise

Build a self-learning agent

Build an agent with skills, MCP connectors, and its own memory and context-retrieval system.

Share your skills in your team's skills library.

Part 1 · Concepts

Context & Memory

How agents stay focused on the right things — and remember what matters across turns.

Why context matters

Building AI systems is all about context

Context window

system

Agent instructions (keep light)

memory

memory.md — what Claude remembers about you

tools

Short descriptions of available tools

user

Your prompt

Too little context → response lacks relevance.
Too much context → window clogs, model loses focus.

Heads-up

Context is a resource you must optimise.

We need context-engineering strategies that put the right information into the window at the right time.

From Session 1

Keep instructions light. Put detailed tool usage (e.g. "best practices for writing an email") inside the tool config, so the agent only loads it when it calls the tool.

Progressive disclosure

A core idea

Progressive disclosure

Don't load everything up-front. Let the agent discover what it needs, turn by turn.

Example · PPT generation

User asks to generate a PPT.

Default context only includes artefact_generator.md.

It tells the model to use ppt_generator.md.

ppt_generator.md loads brand & code modules on demand.

word_generator.md never referenced — saves context.

Keep a large library of skills; let the agent progressively discover the right pieces per task.

Loaded per turn

Turn 1

User prompt + short description of all tools

Turn 2

artefact_generator.md

Turn 3

ppt_generator.md

word_generator.md never referenced

Turn 4

brand_guidelines.md

ppt_generator.py

We will leverage this concept inside Claude Code with the /skills system.

A core idea

Memory — what carries across turns

When memory is blank

Turn 1

User prompt

LLM

Response

Turn 2

User prompt

LLM

Response +
save_to_memory

When memory is updated with previous conversation

Turn 1

memory +
User prompt

LLM

Response

Memory is prepended to the prompt — context flows from one turn into the next.

How it works

Most LLM systems load memory into every prompt.

They have built-in mechanisms to identify and update memory with relevant context — and users can also explicitly ask for things to be added.

Two paths in

Agent decides on its own (save_to_memory tool)

You ask: "remember this…"

Part 2 · Walkthrough

Introduction to
Claude Code

The CLI we'll use for the rest of the workshop.

Walkthrough · 01

What is Claude Code?

Definition

A CLI-based experience for connecting to Anthropic models.

Has access to your local filesystem.

Builds context graphs on your filesystem via progressive disclosure — it keeps learning and getting better.

Quick PowerShell tips

cd [folder] — navigate into a folder

cd .. — back to the parent folder

ls — list files in current folder

mkdir [name] — create a new folder

CLI

① Open

a folder

② Make

a folder

③ Launch

Claude

④ Run

a prompt

Walkthrough · 02

Options available within Claude Code

Typing / opens a dropdown — use ↑↓ to navigate.

/ menu

Useful stuff

/clear — clears your context window

/compact — summarises current conversation to save space

/resume — continue a previous conversation

/exit — quit Claude Code

Will be used in today's session

/mcp — status of installed MCPs

/skills — list all available skills

/agents — create & manage agents

/tasks — see running task processes

/model — change the default model

/config — change overall settings

Esc — stops the current turn

Walkthrough · 03

Other ways to access Claude Code

Claude Code within VS Code

Available to all

Install the Claude Code extension within VS Code — useful to navigate folders and see files created inside the VS Code UI.

Install the Claude Code extension within VS Code.

Claude App

Restricted

Navigate to the Code tab within the Claude app. Currently not accessible for some enterprise tenant logins.

Navigate to the Code tab within the Claude app.

⚠

Important: Using the Opus model within VS Code is not the same as using the Claude Code extension within VS Code.

Walkthrough · 04

Claude Code's core purpose: agentic coding

Everything you learned in Module 1 (vibe coding) applies directly here.

Nebula tip

Exercise segments that require running a prompt are highlighted.

Try this out

01

Start Claude Code in an empty folder.

02

Run the prompt on the right.

Prompt

        Make a hello world app with a working front end and back end

Walkthrough · 05

Context is Claude's superpower

claude.md

Loaded for every prompt

Lives in your root directory. Or just ask Claude to show it to you.

memory.md

Everything Claude remembers about you

Whatever you've asked Claude to remember — attached to every prompt.

Plus

Claude further manages context with additional strategies:

Progressive disclosure via the skills system

MCP tool wiring for external services

Sub-agents with isolated context windows

Compaction to keep long sessions alive

We'll deep dive into each of these.

Walkthrough · 06

A primer on key terminology

Four capabilities you'll touch today.

Claude capability	Definition	Today's exercise
Tool MCPs	LLM protocol to connect with other products/services and perform actions.	WorkIQ — fetch context from your work-graph (Microsoft Graph for M365; similar exists for Google Workspace). ADE MCP — explore your data warehouse (Kusto here; Snowflake / BigQuery / Databricks work the same way), run .KQL / SQL queries.
Context Skills	Re-usable context files (.md) that Claude can refer to when required.	Kusto-skill — a system to query insights from your warehouse with best practices and business rules baked in.
Memory memory.md	Everything Claude remembers about you. Attached to every prompt.	Personal context — key info about you and your work.
Agent Sub-agents	Claude can spawn multiple independent agents with their own context window.	Out of scope Skills subsume most of the advantages of agents — except their dedicated context window.

Part 3 · Hands-on

Session 2
Learning exercise

Memory, MCPs, and skills — wired into your Claude Code.

Exercise · 01 · Memory

Let's leverage Claude's memory

Let Claude know a little about you and the work you do.

What to do

Enter 4–5 important things you want Claude to know about you.

Things like your name, role, charter, team, and how you work.

From now on, Claude attaches these to every prompt automatically.

Once it's in memory.md, you never have to repeat yourself again.

Sample prompt
Remember the following in your memory:

      My name is [Placeholder] and I drive the [placeholder] charter.

      I work with [placeholders] and report to [placeholder].

      My …

Tip

Replace placeholders with your real details. Keep it concise — memory is loaded into every prompt.

Exercise · 02 · MCPs

MCPs extend Claude's context

MCPs connect Claude to external services — Notion, ADO, Kusto, Figma, etc. Use the connection to send/fetch context and invoke tools.

Let's explore the WorkIQ MCP

WorkIQ gives you read-only access to your work-graph — Teams messages, emails, people details.

WorkIQ MCP (Microsoft 365 work-graph); similar MCPs exist for Google Workspace and other ecosystems.

Four steps

01

Install WorkIQ

Prompt: Read about this mcp and install it <your MCP provider's documentation>

02

Accept the EULA

Approve the MCP's terms when prompted.

03

Check it's installed

Prompt: /mcp

04

Run it

Prompt: Use WorkIQ and give me a summary of my emails in the last 24 hours

Exercise · 03 · Skills

Skills help Claude build knowledge on how to do something

Central metaphor

In a kitchen, MCPs are the tools (knives etc.) and Skills are the recipes.

What a skill is

A set of instructions that teaches Claude how to handle specific tasks or workflows.

Instead of re-explaining your preferences, processes, and domain expertise in every conversation — teach Claude once, benefit every time.

When skills shine

Skills are powerful when you have repeat tasks — the same query you run every Monday, the same brand guidelines you apply, the same review checklist you walk through.

Repeat

Standardise

Share

Best practice

Pair your MCPs with skills on how to use them

Same MCP, very different result — depending on whether Claude has the skill that tells it what your "important" looks like.

Without a skill for WorkIQ

Need to ask for what you need in detail.

If you ask "what's the latest messages on Teams?" — Claude doesn't know it should prioritise messages from your managers / LTs or specific important projects (unless you explicitly mention it).

        "what's the latest messages on teams"

        # → flat list, ranked by recency only

With a skill for WorkIQ

Claude knows your priorities and surfaces what matters.

For the same prompt, Claude prioritises important and relevant messages based on how you've taught the skill to rank them.

        "what's the latest messages on teams"

        # → manager + LT + Nebula thread first,

        # → rest collapsed by topic

MCP = the tool · Skill = the recipe for how you use it

Hands-on · Skill #1

Let's build your first skill on WorkIQ

Teach Claude once. Benefit every time you ask for a Teams debrief.

Prompt to Claude Code

Create a "teams-debrief" skill that fetches all messages in the last 24 hours on Teams and gives a summary.

Additionally, surface up messages from my manager along with all of my manager's direct reportees. Also look for any relevant messages on Nebula.

Claude generates the skill

teams-debrief · SKILL.md

Skills that get better · Concept

Building a data analytics skill

Kusto used as the working example — same workflow applies to Snowflake, BigQuery, Databricks.

① Use a knowledge base

Have instructions in your skill to update a knowledge base every time it learns something.

② Save commonly used patterns

Have instructions to save query patterns that work well — so Claude doesn't rewrite them.

Updated by

i) User sharing learnings · ii) Claude learning by itself

Example · Kusto / ADE

Your knowledge base can include:

Business rules

Table schemas and table details

Cluster / connection details

Common mistakes to avoid

Commonly used patterns:

Re-usable .kql / .sql queries indexed for fast retrieval — so Claude doesn't rewrite the same metric query twice.

Skills that get better · Steps 1–3

Building a data analytics skill

Wire the MCP and CLI before you write a single line of skill logic.

01

Install Azure Data Explorer MCP

Or the equivalent MCP for your warehouse — Snowflake MCP, BigQuery MCP, Databricks MCP.

            Prompt: I need to set up ADE MCP to connect to Kusto. Read this documentation and set up the MCP: [your data warehouse MCP docs]
          

02

Install Azure CLI and complete login

Or the equivalent CLI for your warehouse — SnowSQL, bq, Databricks CLI.

Prompt: Set up Azure CLI and complete the login

03

Exit Claude and close your terminal. Open a new terminal and start Claude again.

Verify

Restart your Claude instance and check installation success by typing

/mcp

You should see your warehouse MCP listed as connected.

Skills that get better · Steps 4–6

Building a data analytics skill

04

Check if your skill is created

Prompt: /skills

05

Share connection & cluster details

            Prompt: Connect to this cluster and help me understand the data available [your warehouse URL]
          

06

Learn from existing queries

            Prompt: I am sharing some queries that are already made. Learn about the tables, business rules and approach for key metrics and update your knowledge base.
          

→ Paste any existing queries

How to find your connection details

① Click "more actions" next to your cluster.

② Click "edit connection" and copy the details.

Your warehouse UI will look similar — the connection-string lookup is the same workflow.

Skills that get better · Steps 7–8

Building a data analytics skill

Now run the skill — and teach it what "good" looks like for your product.

07

Run your skill

Prompt: Fetch the latest insights of [placeholder]

Paste your product-specific details

08

Improve your skill

            Prompt: Improve the skill by ensuring to always include production ring and split data across iOS and Android.
          

Rewrite based on what you find poor

Feedback loop

Every "improve the skill" prompt updates the knowledge base. The skill compounds: better next run, better still the run after.

Custom tools · Skill #3

Build your own custom tools

MCPs aren't the only way to invoke actions. Build local scripts or app-specific tools and wrap them as skills.

Prompt 1 · Create the skill

Create a skill that generates professionally formatted enterprise PPTs.

Prompt 2 · Use it

Create a deck assessing the strategy of [your example company] in AI.

Claude's first draft

.pptx

Custom tools · Iterate

Build your own custom tools

The first draft won't be perfect. Give feedback so Claude can learn and do it better next time.

Prompt 1 · Improve the skill

        Improve the PPT creator skill to ensure to include a logo on the bottom right. Follow your org's brand guidelines. Include key takeaways on each slide.
      

Prompt 2 · Run it again

Regenerate the deck.

Improved draft

v2

Custom tools · Skill #4 · Email

Build skills that act on other applications

Pre-req: Classic Outlook (or your email client — Gmail API, AppleScript for Mail, etc.)

Prompt to Claude Code

        Create a skill to send emails to my colleagues using Outlook COM automation. Ensure the email is richly formatted and fetch email IDs from WorkIQ.

        Remember to sign off emails with "Claude on behalf of __________".

Review your email

This was the first draft Claude Code made for the Nebula teaser — pulled from your team's meeting contexts.

Plenty of newsletter-automation opportunities once the skill is live.

Generated email

Outlook draft

The aha moment

Putting it all together

One prompt. Four skills. Two MCPs. Memory and work-graph routing the rest.

Master prompt

Fetch the latest performance data for my product and create a deck with insights, and prepare a draft to review on Outlook to send to my manager.

Invokes

MCP

Kusto MCP

SKILL

Kusto skill

SKILL

PPT generation skill

SKILL

Email skill

Claude Code

Orchestrates

Picks the right skills + MCPs.
Routes context. Sequences calls.

Progressive disclosure — only loads what each step needs.

Resolves

My product

Identified from your memory.

My manager · name

From memory or WorkIQ.

My manager · email

Resolved from WorkIQ.

One prompt → orchestrated workflow → reviewed draft in Outlook

An aside

But what about agents?

Before skills existed, all context loading clogged the main agent's window. Skills subsume most of the benefit — but agents still earn their keep in one case.

Context

Claude lets you create custom agents to solve specific problems.

Prior to launching skills, all context loading into the main agent clogged the context window. Agents fixed this by giving each task its own instructions and context.

With the launch of skills, most benefits of building agents separately don't exist — skills cover the same ground with less ceremony.

Where agents still help

Each agent gets its own context window.

For large tasks — e.g. analysing eval test cases — spawn multiple agents so each has its own clean window to work in.

Prompt: Make multiple agents and complete this task…

Map-reduce style: parallel agents, each with their own context, results aggregated by the main thread.

Scope

We will not cover making agents with specific context/instructions — building skills subsumes that.

Session 2 · Recap

Learning recap

Concepts & foundations

Understood how to use context & memory along with their implementation in Claude Code.

Learnt how to navigate the Claude Code CLI interface.

Built a hello world app using Claude Code.

Learnt about MCPs and connected to WorkIQ and Kusto MCPs to fetch external data.

Skills & custom tools

Built skills (progressive disclosure) to ensure Claude has just the right context when calling MCPs.

Built learning systems using skills to make Claude get better every interaction.

Made a skill that leverages custom scripts to generate PowerPoint decks.

Made a skill that leverages other applications on your local — sending email via Outlook.

Next up

In the next session we'll learn how to orchestrate these skills to build your own workOS.

Project Nebula · Module 3 · Session 2

See you soon!