AI bootcamp for EU affairs, day 10: build the agent, block by block
One agent, five layers. Start with a sentence, finish with an assistant that explains the AI Act to four different audiences and tells you when it is unsure.
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
A full worked build of an AI Act explainer agent, in five cumulative instruction layers with a test prompt for each, followed by a blueprint for designing your own.
Who it is for
EU public affairs, policy and communications people who have built a first project or agent and want to know what separates a good instruction set from a long one.
When to use it
Use it as the hands-on session after day 9. Day 9 gets you a project that behaves; this makes it genuinely useful on a file where accuracy matters.
Key takeaway
A good agent is assembled, tested after each addition, rather than written in one go. Five layers, five tests, and you can tell exactly which instruction produced which behaviour.
In brief
- Build one agent in five layers, testing after each, so you can see what every instruction actually changes.
- The subject is the AI Act, because it is the file where a confident wrong answer costs the most.
- Layer 1 tailors the audience. Layer 2 makes it ask before it answers. Layer 3 gives it two roles that check each other.
- Layer 4 stops it inventing law. Layer 5 grounds it in the official text you upload.
- Then design your own, using the blueprint at the end. The exercise matters more than the example.
Framework
Five layers, five tests
Each layer adds to the one before it. Test after every layer, because the point is to see the change.
| Layer | Label | Description |
|---|---|---|
| Layer 1 | Customise | Role, subject and audience. One sentence that decides everything else. |
| Layer 2 | Flip it | It asks a clarifying question and proposes a plan before it answers. |
| Layer 3 | Persona | Two roles in sequence: a communicator who explains, a reviewer who flags risk. |
| Layer 4 | No invention | Fact-checking rules, and permission to say it does not know. |
| Layer 5 | Ground it | The official text uploaded, and rules for quoting and labelling sources. |
Five layers, and a test after each one
Day 9 left you with a project that behaves consistently. This is the session where it becomes genuinely useful, and the method is the opposite of what most people do.
Most people write one enormous instruction set, paste it in, and hope. When the output is wrong they add another paragraph, and another, until nobody can say which line is doing the work and which line is making things worse.
Build it in layers instead. Five instructions, added one at a time, with a test prompt after each one. You will see exactly what each layer changes, which means you can debug it later and explain it to a colleague.
The subject is the Artificial Intelligence Act, chosen deliberately. It is long, it is contested, guidance is still arriving, and a confident wrong answer about it can embarrass an organisation in public. If the method holds here, it holds on your files.
By the end you will have an assistant that explains the AI Act to an MEP, a journalist, a startup or a lawyer, asks what you actually need before it writes, separates what the law says from what it means, and tells you plainly when the answer is not settled.
Every prompt from this bootcamp is collected in the AI prompt library.