Series: How I built a second brain · part 4
I have six AI advisors, and each of them sees only its own field. That's how I designed them, and that's how it should be. But almost every question that really matters sits on the border between fields. The business coach says "push now", the health coach says "you're running on reserves", and both are right. So on 15 June I sat them down at one table.
The council is not a seventh advisor
The first decision was what the council is not. It is not a new advisor with opinions of its own. It is a format: a moderated round table of the existing six voices. The life coach chairs it, because it already held the whole picture while the others dealt with the "what" and the "how".
The conversation has four steps:
- Framing. The moderator restates the question once, in a sentence or two, and checks what is really being decided. If the real question is different from the one asked, it says so straight away.
- Opening positions. Everyone gets the floor. An advisor with nothing to say on the topic says so in one sentence and doesn't fill the space.
- Clash. Only where two advisors genuinely disagree. Where they agree, the moderator notes it and moves on. It looks for real fault lines, not theatre.
- Agreement and dissent. One recommendation the council stands behind, and right below it the minority view and the tension it came from.
The council must not average. Six opinions averaged into one are mush; six opinions with their disagreements named are a map.
The other rules followed from the same idea. Each advisor speaks only within its own field: on a tax question the financial advisor decides, not the moderator. The moderator connects; it doesn't overrule. And every session ends with what is left to me. The council recommends; it doesn't decide.
I solved memory as simply as I could. A council session is one event, so it gets one entry in the shared log, not six. The council has no memory page of its own. Anything lasting I learn from it goes into the shared profile or onto the page of the advisor it concerns.
When the council, when one advisor
I call the council only when a question crosses fields or involves a real trade-off: should I take this on? What am I giving up for it? I'm being pulled in five directions and can't see the whole. A tax return deadline or a change to my training belongs to one advisor. For those, the council is expensive and pointless.
From conversation to morning meeting
For two months I convened the council myself. Meanwhile scheduled checks were running, one for each advisor. By August there were twelve of them, and each one messaged me separately.
On 19 August I replaced them with a single morning meeting. The members first prepare separately: each reads only its own page and the morning summary of mail and files, returns at most three lines, and may say it has nothing. Only the council sees all six at once, and it sends me one recommendation.
This also got rid of one specific bug. The old check asked me every morning what I would finish that day. It had nowhere to store my answer and no notion of "nothing", so it asked the same thing three mornings in a row. A question whose answer the system has nowhere to store will be asked forever.
Four hours instead of advice
On 24 August I gave the council four hours of my day, from noon to four. The plan for those hours replaces the daily recommendation; it isn't added to it.
The meeting stayed the same. Only the output changed, and it is harder. A recommendation can be one good idea. A plan has to survive the hours: every task estimated, each with a condition for when it is done, ordered so that the first doesn't depend on the third. The last half hour stays free, because every plan overruns. Whatever is already in the calendar comes first. And a quiet day gets protected free time, with one sentence saying the council found nothing urgent. It doesn't invent work.
What went wrong
The block nobody opened
At first the plan went into the calendar as one event covering all four hours. That was safe: one event with one fixed name can't make a mess. But in the calendar I saw an opaque bar, and the individual tasks only when I opened it. A plan you have to open is a plan you don't act on. Since 28 August every task has been its own event.
A permission that couldn't be narrowed
The council was allowed to edit events too, and for a short while to delete them. I took both away the same day. The calendar tools couldn't be restricted to one calendar: for MCP tools, the list of allowed tools doesn't restrict them at all. That it could touch only its own calendar was a sentence in the prompt. A sentence in a prompt is not a restriction. A tool that can rewrite a meeting with another person has to carry its boundary inside itself.
Adding is a different kind of permission, not a smaller dose of the same one. A wrongly added event is visible, and I can delete it with two taps. Today the council can only add, and a test makes sure the permission doesn't quietly widen when code is copied. I pay for this: the council can't correct what it has written itself. When it makes a mistake, it says so in its message and leaves the event where it is.
A limit I never asked for
In early September I noticed in the calendar that the council never planned more than three tasks. I had never wanted that limit. It was one sentence in the prompt, "two or three segments, not four or five", written at the same time as the four-hour block and repeated three more times. A cap on the number of tasks is a cap on the plan: the estimates bend to fit the number, instead of the number coming out of the estimates. Today each task gets as much time as it needs, and as many go in as will fit.
The council that assigned me work
On 1 September the council scheduled into my block a piece of research that an agent could have done. I told it I didn't want it to tell me what to do; I wanted it to do what it could do itself. Since then only my own work goes into my block, mostly judgement on what the agents have prepared. How the council got hands is the subject of part nine.
What I learned
A council makes sense where fields meet, because those seams are exactly where individual advisors go wrong. Its value lies in named disagreement, not in agreement.
And every extension of authority has to be narrowed where code enforces it, not text. All three calendar mistakes had the same root: a rule or a limit lived only in the prompt, where nobody saw it.
How to start
- Don't make the council another advisor. Make it a format for the ones you already have.
- Give it a moderator who connects but doesn't overrule an expert in their own field.
- Require agreement and dissent separately. Never just the result.
- Let the members prepare separately and see each other only at the table.
- If the council may act, let it only add things at first. What can't be restricted in the tool, don't grant at all.
- Read the prompt as code. Every sentence in it is a limit, even the one you wrote in passing.
Next time: how I moved the advisors to Telegram.