Writing what must survive, deciding without theatre

A conversation is an excellent tool. It lets you quickly explore a problem, understand a nuance, resolve a disagreement or simply work with other people.

But a conversation has an obvious weakness. It disappears.

Those who were present remember it differently. Those who were not know nothing about it. A few months later, a new person joins the team and has to rebuild the reasoning from fragments.

In a company distributed across several countries and time zones, that limit becomes particularly costly. So we have a simple rule:

What must outlive the conversation must leave the conversation. An important decision becomes a written artifact. An important lesson becomes a reusable document. A procedure joins the reference base. The rest can disappear.

We are not trying to document everything. A company where every sentence must be recorded would quickly become as bureaucratic as the one we are trying to avoid. We try instead to document what will let someone work tomorrow without needing to track down the right person.

An important decision is written down

At Cercle, a durable decision has an owner, a context and a written trace. For important topics we use a format inspired by RFDs, Requests for Discussion.

The idea is not to write a perfect report before starting to think. It is exactly the opposite. The document appears early. It can be incomplete. Others can comment, challenge an assumption, add a constraint or propose an option. The reasoning becomes visible while it is being built.

The method has an important advantage. A decision does not only keep its result. It also keeps its history.

Two years from now, someone will be able to understand why an option was set aside. And if the constraints have changed, that person will be able to challenge the decision on better grounds than “I do not understand why we did this”.

Eliminating before choosing

For structuring decisions, a single option presented from the start is often a bad signal. It sometimes means the decision is already made and the document mainly serves to justify it.

We open several possibilities, we test them when possible, we define criteria, then we progressively eliminate what does not hold. A structuring decision document that presents only one option is sent back to its author.

Take three technical solutions. The first exceeds the acceptable latency. It is eliminated. The second costs three times as much to operate. It is eliminated. The third remains.

The final decision may fit in one sentence. But the two eliminations make up a large part of the knowledge produced.

This way of deciding also reduces the influence of status. The option proposed by the most experienced person is not automatically chosen because it comes from the most experienced person. Experience helps to form better hypotheses. It does not turn a hypothesis into a fact.

One person decides, several people can contradict

We reject permanent consensus. We equally reject arbitrary decisions.

Every topic therefore has an owner. That owner must gather the relevant information and listen to the people concerned. Then they decide.

The disagreement stays visible. A person can have argued seriously for another option and then contribute fully to the final decision.

We consider that distinction essential. A healthy system must be able to carry a lot of disagreement without becoming unable to act.

Messaging stays ephemeral

Messaging tools easily give the impression that documentation is solved. Everything is written down somewhere. But “somewhere” quickly becomes a problem.

A decision buried in six months of conversations is not really accessible. It is neither structured, nor versioned, nor maintained.

We therefore treat messaging as an ephemeral tool. It is for talking. Not for holding the company’s official memory. What matters must leave it.

Reusing what we have learned

An organisation can accumulate an enormous amount of experience and keep repeating the same mistakes. All it takes is for each lesson to stay in the head of the person who lived it. We want to turn part of that experience into collective memory.

Useful lessons become short briefs, written while the work is still fresh. Their title already states what was learned.

Not “thoughts on agents”. Rather “beyond twelve steps, our agents drift without a checkpoint”. The first announces a topic. The second already conveys information.

Essential knowledge can then be turned into short training materials and presented again regularly. The goal is not to build the largest possible wiki. It is to make the organisation progressively more competent.

AI makes this discipline even more important

Agents now let us produce far faster than we can verify. That is a considerable opportunity. It is also a new problem.

A generated summary can invent a decision. A document can mix verified information with supposition. An error can be picked up by another agent and propagate.

Generated but unreviewed content remains a draft. Validated content has a responsible human. Only canonical information is authoritative.

Agents can propose, draft, collect and prepare. They can even take part in the workflow. Responsibility remains human.

In the end, this distinction offers everyone something simple. You should be able to go on holiday without becoming your team’s human DNS server. And when you join a topic you do not yet know, you should be able to understand why things are the way they are without having to reconstruct several years of private conversations.