Workshop guide

This page has everything from the live session: the definitions, the account set-up, the step-by-step build, and a Copilot equivalent. If you followed along live, this is a record to come back to. If you missed it or want to run it again at home, you can work through it top to bottom. The downloadable Project instructions and Skills themselves live on the resources page - this page is about how to use them.

Session agenda

11:30	Acknowledgement of Country, introductions
11:35	Context and definitions - what's an assistant, anyway?
11:40	Set up - Claude account, this website
11:45	Step-by-step: building the AI assistant
12:15	Troubleshooting and bonus steps
12:30	End

Assistant, or agent?

Both words get used loosely, and it's worth being precise before you build either one. A chatbot is a conversation - minimal configuration, starting fresh every time. An AI assistant is a configured conversation with standing instructions and your own material attached, so it answers the same way every time and can reach into your other apps - but only ever with your say-so, right at that moment. An agent is the same capability with that approval gate removed: it acts on its own, on a schedule or a trigger, with nobody checking each step.

	Basic chatbot	AI assistant	Agent
Configuration	Minimal, starts fresh each time	Persistent instructions + your own material	Assistant, plus permission to act
Consistency	Depends how you phrase it, every time	Same rules applied every time	Same, plus a record of what it did
Grounded in your work	No - general knowledge only	Yes - your context, your data, your examples	Yes
Takes actions in other apps	No, only talks	Yes, but only with your approval at the moment it happens	Yes, without needing approval each time

Today we're focused on assistants, and we recommend following the human-in-the-loop (HITL) principle throughout: check each step and action, especially while you're still building a flow.

One other pro of assistant is reproducibility and being able to trace back to the source. A chatbot fails both tests: ask it to code the same transcript twice and you may get two different answers, with no fixed rule to point to when someone asks "why did you code it that way?" An assistant fixes both problems at once: the same codebook every time, and an actual instruction you can show someone.

What we'll do today

However you're joining - following along on your own computer, going at your own pace, or simply listening - by the end of this session you'll have seen a working assistant built using Claude Projects and Skills. We'll build it against a live example: Burrow Watch, tying the exercise back to evaluation.

Build an AI assistant in Claude for the Burrow Watch NGO
Explore how to set up your environment for quality and security
Use Skills and Connectors to create optimised, reproducible tasks
Explore bonus steps and alternatives (e.g. Copilot)

We'll aim to keep the last 15 minutes for questions, troubleshooting, and bonus steps.

Setting up

Screenshot: Claude account settings — Claude account settings - profile, privacy, and model choice

Claude account - go to claude.ai and set up a free account if you don't have one yet. A free account gives you enough credits to build this assistant in the session, but credits are limited, so be mindful of how many times you re-run big steps.

This website - keep burrow.watch open as your go-to for context about the NGO and anything you need to download. You can come back here after the session and follow the same steps at home.

Your profile - in Claude's settings, add a line about who you are (e.g. "I'm an evaluator and I like concise outputs."). This helps every conversation start with useful context about you, not just the Project.

Privacy - check your account's data settings and confirm your conversations aren't being used to train models, especially before uploading anything with real program or client data.

Choosing a model - different models trade off speed against depth of reasoning. For coding transcripts or drafting a report, a stronger reasoning model is usually worth the extra time; for quick lookups, a faster model is fine.

Connecting to the web - Claude can search the web live when a task needs current information (like this quarter's conservation news). Make sure web search is switched on for that step.

Key concepts: Projects, Skills, Connectors, Artifacts

Screenshot: a Skill and a Connector attached to a Claude Project — A Skill and a Connector attached to a Claude Project

Projects

A Project is a persistent workspace in Claude: standing instructions and attached knowledge files live here, so every conversation you start inside it begins with the same context instead of you re-explaining or re-uploading it each time. Everything in this workshop happens inside one Project - it's the container the Project instructions, knowledge files, and all six Skills below sit in.

Skills

A Skill is a saved, reusable set of instructions for a specific kind of task - a codebook to apply, a report format to follow, a search strategy to run. Instead of re-explaining "code this against our six categories" every time, you attach the Skill once and Claude follows it consistently. Burrow Watch's Project uses six: email intake, transcript coding, context research, report generation, report sending, and stakeholder briefs.

Connectors

A Connector links Claude to another app - like your email inbox or Canva - so it can read from (or, with your approval, act in) that app directly instead of you copying and pasting content back and forth. Connectors are what let a Skill like "email intake" actually reach your inbox. Be careful what you connect: only connect accounts you're comfortable Claude reading from, and check what each connector can and can't do before you rely on it.

Artifacts

An Artifact is a standalone piece of output Claude creates alongside the conversation - a document, a webpage, a chart - that you can view, edit, and keep separately rather than scrolling back through chat to find it. We'll use one to build the "State of the Burrows" report itself, styled to match this website.

Building the assistant, step by step

These are the steps we'll work through live. Each one builds on the last, so it's worth doing them in order the first time.

Step 1

Give the assistant context. Download the theory of change, the Project instructions, and the program background from the resources page and attach them to a new Claude Project. This grounding is what lets the assistant give useful, specific answers instead of generic ones.

Screenshot: attaching knowledge files to a Claude Project — Attaching knowledge files to a new Claude Project

Suggested prompt

Copied

I've attached our program background, theory of change, and Project instructions. Please confirm you've read them, then summarise the six coding categories and the five theory-of-change outcomes back to me in a few bullet points.

Step 2

Quantitative data analysis. Upload the zone and funding CSVs and ask Claude to summarise the quarter's trends. Check the result before trusting it.

Screenshot: a quantitative data summary — A quarter's zone and funding data, summarised

Suggested prompt

Copied

Here's this quarter's zone and funding data. Summarise the key trends by zone and month, and flag anything that looks concerning or inconsistent.

Step 3

Additional scan. Turn on web search and run the context researcher Skill to pull in current wombat conservation news, so the report reflects what's happening in the field right now, not just inside the program.

Screenshot: a current wombat conservation news brief — A live external context brief

Suggested prompt

Copied

Search the web for current news on sarcoptic mange in wombats and wombat conservation in Australia. Summarise anything relevant to our program in a short brief, with sources.

Step 4

Transcript analysis. Upload the transcript coder Skill and an interview transcript, and ask Claude to code it. Look at the result together - does it match the framework, and would you trust it as a first pass?

Screenshot: a coded transcript output — A transcript coded against the six-category framework

Suggested prompt

Copied

Here's an interview transcript. Code it against our six-category framework, quoting the relevant line for each theme you identify, and note anything that doesn't fit a category cleanly.

Step 5

Looking at emails. Connect Claude to an email inbox, and use the email intake Skill to find and extract relevant field notes - filtered by subject or sender, never a general inbox browse. Confirm which emails to include before anything is coded (this is the HITL principle in action), then hand the extracted text to the transcript coder. No live inbox handy? Use the sample field-note emails instead - paste them in and the Skill works the same way.

Screenshot: connecting an email inbox and extracting field notes — Extracting field notes from a connected inbox

Suggested prompt

Copied

Check my email for anything with "field notes" in the subject from the last month (or: here are three field-note emails, pasted below). List what you find first, and wait for me to confirm which ones to use before coding anything.

Step 6

Drafting, then building, the report. We recommend a human takes the first pass at writing the report - AI is most useful here as an additional layer of analysis, pulling the coded themes, data, and context brief together. Once you're happy with the content, ask Claude to build it as an Artifact, following this website's design as closely as possible.

Screenshot: the State of the Burrows report as an Artifact — The report, built as a styled Artifact

Suggested prompt

Copied

Using the coded transcript themes, the data summary, and the news brief, build the State of the Burrows report as an artifact, styled to match this website. For each theory-of-change outcome, state on track, at risk, or insufficient data, with a one-line reason.

Bonus steps

If there's time left, or you want to keep exploring at home:

Sending it

Sending the report. Using the same email connector from Step 5, ask Claude to send the report. The report-sender Skill always drafts and waits for your go-ahead before it actually sends, so treat this as safe to try.

Screenshot: a drafted report email awaiting approval — A drafted email, waiting for approval before it sends

Suggested prompt

Copied

Draft an email to [name/address] with the report attached. Show me the draft and don't send it until I confirm.

Meeting prep

Ask Claude to brief you before a call with one of the stakeholders, using the stakeholder brief Skill.

Canva connector

Connect your Canva account so Claude can help design materials directly, rather than only producing text.

Using this in Microsoft Copilot instead

Everything above is written for a Claude Project, but the underlying idea - instructions plus grounding material, applied consistently - carries over to Microsoft Copilot's Agent Builder. A few things to adjust along the way:

Instructions and knowledge: paste the Project instructions text into Agent Builder's instructions field the same way, and attach the background, transcripts, theory of change, and data files as knowledge sources.

No separate skill files: Agent Builder doesn't have an equivalent to individually installable Skills. Fold each Skill's instructions into the main instructions field as separate labelled sections instead of six separate files.

Email intake: needs the Outlook/Microsoft 365 connector rather than Gmail. Read access works the same way; check current Copilot documentation for exact setup steps, since connector details change over time.

Sending the report: Agent Builder in Microsoft can't take actions like sending mail on its own; that needs Copilot Studio's Actions layer instead. Plan to draft the email manually from Agent Builder's output, or build the sending step separately in Copilot Studio.

Running multiple prompts: Agent Builder tends to work better with one instruction at a time rather than a long chained request. Try breaking the full-pipeline prompt above into its separate steps if it doesn't work as a single paste.