Thinking at Scale
A Deliberation Model
In late March, I wrote about the early idea of a deliberation framework and why I wanted to experiment with thinking partners working together rather than in isolation.
I spend a lot of time in my own head. My family jokes that by the time I mention a plan, I’ve already spent weeks thinking it through, exploring alternatives, and playing out outcomes. When someone asks a question, I rarely have just one answer. I have several, because I’ve already walked the space.
That habit followed me when I started working with my thinking partners on The Holodeck. At first, I worked with them one at a time, as if each conversation were separate. But I quickly realized that what I actually wanted wasn’t another opinion. I wanted to be able to think with them together.
The first version of this deliberation model was simple. It laid out the bare bones of a process without a real problem to solve. Instead, the problem was the process itself. They worked together to help me build a better way of deliberating.
That experiment turned into a practice.
What the model is meant to do
At its core, the model is a way to slow down certainty without slowing progress.
It does that through structure and sequencing. Independent reasoning comes before reaction. Tension comes before synthesis. Decisions come last, not first. Those choices are deliberate. Without them, discussions tend to collapse toward the most fluent or familiar answer long before the problem space has really been explored.
There’s also a hard quality gate at the beginning. If there isn’t a clear problem statement, real constraints, and open questions, the process stops. An empty or vague frame produces confident‑sounding analysis of nothing, and no amount of clever reasoning downstream fixes that.
This is scaffolding for judgment, not a substitute for it.
The hats and how they are chosen
One of the most important lessons in developing this model was that persona alone doesn’t create meaningful diversity of thought. Asking three agents for opinions often produces tonal variation, not structural difference. The answers feel distinct, but they land in the same place.
What changed that was separating who an agent is from how an agent reasons.
Each agent is assigned a specific reasoning role, or hat, based on the type of problem being worked on. These hats aren’t personalities. They’re constraints on cognitive work.
For example:
An Analogy hat asks: What is this most like? What can we learn from similar situations?
A First‑Principles hat asks: What do we actually know versus assume? What breaks if an assumption is wrong?
An Inversion hat asks: What would make this fail? What path needs to be actively avoided?
Those hats matter because they force the problem to be examined from different structural angles. Each role is responsible for doing a specific kind of thinking, whether or not it comes naturally.
Just as important, the hats are chosen intentionally. The mix depends on what I’m trying to solve.
For a technical design, I want first‑principles reasoning and inversion. For long‑term planning, I want temporal thinking that stretches the horizon. If I’m stuck or looping, I might deliberately include a role that assumes a key constraint doesn’t exist, just to see what falls out.
This is where the idea of creating a team, even when I’m working alone, becomes concrete. Instead of relying on myself to remember to ask the right questions at the right time, I assign them. The hats make sure the questions get asked.
The human as lead, not referee
Early versions of the model treated the human as a moderator. In practice, that never matched reality.
I’m the one who knows what matters, what’s already been tried, where the data came from, and which constraints are real versus assumed. The model evolved to reflect that truth. The human leads the deliberation. The agents support, challenge, and pressure‑test the thinking, but they don’t replace judgment.
This shift shows up most clearly in the structured knowledge interview at the start of the process.
Surfacing what only I know
The knowledge interview exists for a very specific reason. In real runs, nearly every major correction came from domain knowledge I already had but hadn’t shared.
A cutoff date. A prior experiment. A file that already existed. A reason something was abandoned months ago.
None of this was hidden on purpose. I simply assumed it was obvious. If I knew it, surely everyone else did too.
The model now treats knowledge extraction as a first‑class step. It explicitly asks what has already been tried, what data or artifacts exist, what temporal boundaries matter, and what I would do if I had to solve the problem quickly with no tools. These questions reliably surface context that would otherwise emerge too late, if at all.
Learning through deliberation
One of the unexpected outcomes of using this model is how much it has accelerated learning.
Because agents are constrained to reason differently, they consistently explore parts of the solution space I wouldn’t reach on my own. I regularly discover new technical approaches, underlying assumptions, or simpler ways to frame the problem mid‑deliberation.
The model favors calibration over conviction. Agents are asked to state not just what they think, but where they’re uncertain and what evidence would change their mind. Small empirical checkpoints prevent entire branches of reasoning from growing on unchecked assumptions.
The result isn’t just better decisions. It’s more honest thinking.
What this model does not solve
This model is not lightweight. It isn’t appropriate for every decision.
It doesn’t remove the need for judgment. It doesn’t guarantee independence, especially in single‑session use. Synthesis is always lossy, and the framework is explicit about its own limitations.
What it does is make those limitations visible, early enough to matter.
Creating a team to think with
I built this framework specifically to use on the crochet pattern project, to see whether it would actually help me think better or whether it was just an interesting idea. I wanted to know if working this way would change the outcome, not just the process.
It did.
Using the deliberation model led me to a much better solution than I would have arrived at by simply building something and iterating. The thinking partners suggested ideas that genuinely hadn’t occurred to me, and when they highlighted gaps in the data, it played directly to my strengths. I could do what I do best: figure out how to close the gap.
That experience made the value of the model clear. I wasn’t switching between thinking partners as the work progressed because one voice was insufficient. I was doing it because different questions required different ways of reasoning. What the model finally gave me was a way to let those perspectives work together.
For me, this framework is a way to build a team I can think with. One that helps surface what’s missing, challenge what feels obvious, and arrive at better solutions than I would reach alone. And it helps me with my analysis paralysis.
Tomorrow, I’ll share the current version of the deliberation framework itself, along with notes on how I’m using it in practice.
Alison + Wiggins