AI bootcamp for EU affairs, day 9: projects before agents
Four platforms, one method. How to build your first working assistant in Microsoft 365 Copilot, Claude, Mistral and ChatGPT, and why the project comes before the agent.
Sebastián Rodríguez Pérez is a European campaign strategist with direct experience running and advising campaigns for some of Europe's largest pro-European organisations. He founded the European Campaign Playbook to codify practitioner knowledge across the EU's 27 political environments and make it accessible to campaign professionals, advocacy teams, and civic organisations working at the European level. His work spans electoral strategy, EU public affairs, digital campaigning, and the integration of AI tools into political communication workflows.
Disclosure
Sebastián Rodríguez Pérez is the founder of the European Campaign Playbook and Campaign Intelligence Library. He has worked with pro-European organisations across multiple EU member states. He maintains editorial non-partisanship in all published content; where articles relate to campaigns or organisations he has directly worked with, this is disclosed within the article. He holds no current positions in active electoral campaigns.
What it covers
How to build a first working AI assistant on Microsoft 365 Copilot, Claude, Mistral and ChatGPT, the difference between a project and an agent, and the same weekly EU monitoring exercise built four ways.
Who it is for
EU public affairs, policy, advocacy and communications professionals who read day 8, understood what an agent is, and now want one that works.
When to use it
Use this when you have decided a task is worth automating and you need to actually build the thing, on whichever platform your organisation already pays for.
Key takeaway
Learn one method rather than four products. Instructions, knowledge, capabilities and boundaries are the same four parts on every platform, and the vendor names change faster than the ideas do.
In brief
- A project is a workspace that remembers. An agent is a worker that acts. Build the first before you build the second.
- All four platforms use the same four parts, under four different names: instructions, knowledge, capabilities and boundaries.
- Microsoft 365 Copilot Agent Builder and ChatGPT Workspace Agents will build you a real agent. Claude Projects, ChatGPT Projects and Mistral Projects give you a context that behaves consistently.
- Build the same weekly EU monitoring assistant on whichever platform you have, and you will understand all of them.
- Give a first agent read access and nothing else. Write permissions can wait until it has earned them.
Framework
The same four parts, four different names
Every platform in this article has these. Learn them once and the interfaces stop mattering.
| Layer | Label | Description |
|---|---|---|
| What job is this doing, and how? | Instructions | Role, mission, workflow, audience, format and rules. Called instructions, project instructions or system prompt depending on who built it. |
| What may it trust? | Knowledge | Approved positions, legislation, briefs, terminology, past work. Called knowledge, files, project knowledge or sources. |
| What may it do? | Capabilities | Search, read files, reach connected systems. Called capabilities, tools, actions, apps or skills. |
| What must it never do? | Boundaries | The list of things it may not touch. Rarely a single screen, usually spread across permissions, approvals and whatever you wrote in the instructions. |
A project remembers. An agent acts.
Day 8 explained what an agent is and what it needs. This one builds it.
There is a mistake almost everyone makes at this point, and it costs a fortnight. You hear the word agent, you open whichever tool your organisation pays for, and you reach for the most autonomous thing on the menu. Then you spend two weeks debugging a system that is doing three jobs badly, and you conclude that the technology is not ready.
The technology is usually fine. The order is wrong. Start with a project.
A project is a workspace that remembers: instructions you wrote once, files you trust, and a running conversation that stays on the subject. An agent is a worker: it takes that same context and acts on it, on a schedule, using tools, without you in the room.
Build the project first and you get two of the four parts right, instructions and knowledge, while the cost of a mistake is still a bad paragraph on your own screen. Add capabilities and autonomy afterwards, once you already know what good output looks like.
What follows is the same first assistant built on the four platforms you are most likely to have: Microsoft 365 Copilot, Claude, Mistral and ChatGPT. The screens differ. The method does not.
Every prompt from this bootcamp is collected in the AI prompt library.