The five-minute test

Does your AI read the manual?

The method is agnostic: it works with the AI you already use, in chat mode or in project mode. But “it works” does not mean “it works equally well in all of them”, and we would rather you knew that before your first session than halfway through it.

We do not give you a table of which model and which plan perform best. It would be stale before you read it, because models change their behaviour every few months. We give you something better: a test that takes five minutes and that you can repeat every time you change AI.

01Two ways of working, a single method

There are two ways of using an AI today, and it is worth knowing which one you are in before loading anything. The method is the same in both. What changes is the mechanics of loading, saving and starting each session. And a new risk that only appears in the second.

Chat mode

One conversation per tool

You open a conversation, attach a document, you talk. When you close it, the conversation stays there and the next one starts from zero. It is what any AI does in its basic version, and what almost everyone working inside an administration has, with the tool they were given.

Project mode

A space that remembers

You create a space with permanent files and instructions, and every new conversation is born knowing what is in it. Many AIs already offer this under different names: projects, spaces, gems, custom agents. Some go further: they read and write files in a folder of yours, search the internet on their own and run multi-step tasks without asking. That is an agent.

02The four minimum capabilities

Do not ask which is the best AI. Ask whether yours has these four things, because they are the ones the method actually uses.

01 · Decisive

Long conversation memory

A diagnosis session is forty minutes and twenty turns. If the model loses the thread halfway and asks you for the territory again, the session breaks. It is the most decisive capability, and the one that varies most between free and paid plans of the same product.

02 · Convenient

File upload

Being able to attach the tool document is far more reliable than pasting it into the chat. If your AI does not take attachments, you can paste it, but paste it whole and in one go, never in pieces.

03 · Solvable

HTML generation

The four objects are delivered as a self-contained HTML page. Any current model knows how to write one; what changes is whether it shows it to you rendered or hands it over as a block of code. If it gives you code, nothing is lost: it takes two minutes to save.

04 · Silent

Stamina in the long session

Some plans cut long conversations short or degrade quality as the context grows. You will notice because the questions turn generic and stop quoting your material.

03The test

Load the document for tool 01 and type exactly this:

What are you going to do with me, and what is your first question?
Sign that it is going well

It answers in a few lines, naming the Collective Intelligence Canvas, announcing that you will work one point at a time, and asks you for four things: territory, promoter, problem and current point. Nothing else. It has read the document as an operating manual.

Sign that it is going badly

It gives you back a summary of the document. Or it offers you a ready-made diagnosis. Or it throws all twelve questions at you at once. Or, if it is an agent, it tells you it is going to “analyse the project” and sets to work without asking you anything. It has not read it as a manual: start a new conversation and load it again. If the second time is the same, try another model or the paid plan of the same one.

If in project mode you have all four manuals loaded, run the test naming the tool: “Today we work on tool 01. What are you going to do with me, and what is your first question?”. If it answers mixing in things from the other three, the space is not separating the documents properly: run the test in a clean chat with only manual 01.

And thirty seconds more, if you load the knowledge base

The package comes with two documents that go up alongside the manual: the Civic Design framework and your project dossier. To know whether they went in, write this before you start:

Before we start: in one sentence, what do you know about my project, and which tool are we working with today?

If it sums up your territory, your promoter and your current point, and names the right tool, the base is loaded. If it answers in general terms about civic design, or names another tool, it has not read the attachments: start a new conversation and upload them again.

04If your AI is an agent

An agent wants to finish the task. The method wants you to get out of your chair between sessions. Out of that clash come three behaviours that hardly ever appear in chat mode and are the norm in project mode. The four manuals carry a written rule against each one, so the AI recognises the correction. Under each symptom, the exact phrase.

Symptom 1

It answers its own questions.

It reads the four lines, deduces the rest and presents you with a finished diagnosis. It looks like efficiency. It is the gate, skipped.

“The answers are mine. Delete the states I have not given you, ask me the first question and wait.”
Symptom 2

It searches the internet for what you have not given it.

You do not know your district's participatory budget; the agent finds it on the council website and puts it into the resources point. It looks like a fact. It is a gap in disguise, because nobody from the territory has told it to you.

“What I have not told you does not exist for this session. Don't search: ask. Remove whatever you brought in from outside and mark it as a gap.”
Symptom 3

It chains the tools.

It closes the diagnosis and, without being asked, opens the actor map “while it's at it”. It looks like a favour. It is a month of territory skipped in a minute.

“One tool per session. Close the object and stop. The next one starts when I come back from the territory.”
The most important one fits in a line, and it is worth learning: “What I have not told you does not exist for this session. Don't search: ask.” An agent that investigates on its own replaces what you know about the territory with what the internet says about the territory. The result looks better and is worth less.

We start where you are. The horizon is local and auditable.

Civic Design defends auditable infrastructures and technological sovereignty; this method leans on the general-purpose AI you already have. It is not a resolved contradiction, it is a visible seam, and it has two practical consequences: do not paste neighbours' personal data or minutes with names into the chat, and keep in mind that what you write travels to a server that neither you nor we control.

Start today · €19 → Back to the method