An AI system can use an earlier decision only if it knows about it. It can use a source only if it can reach it. What happens before the answer is therefore as important as the answer. We are interested in how existing material becomes something useful. That is a job of its own, with its own choices: selection, accuracy, access and later reuse.
The task decides the selection
We call this material context. It includes documents, but also the request itself. It includes permitted actions, earlier decisions and open questions. Context is more than text in an input box. It describes the situation in which a system should understand something or do something. A good model and good context work together. Neither replaces the other.
An example makes it concrete. An AI is asked to explain how an app signs people in. For an overview it needs the parts involved and how they fit together. If it then has to fix a bug, it needs the exact places in the code and the matching tests. It is the same file. But the right view is a different one.
That is what the reading views in LeanCTX are for. The documentation describes an overview of the structure, the full file, selected lines and the changes since the last read. The aim is a view that fits the task. Someone checking a change closely needs different detail from someone getting oriented. [1]
Every selection remains a guess about what matters. So the way back to the original has to stay open. A short view must not become a dead end. If a question comes up about an exception or a missing passage, the system needs to look closer. We make information available step by step, without giving up accuracy.
The right information is the information that makes the next step work.
From choosing to keeping
For us, Context Intelligence starts with a decision. What is the task, and what could matter for it? Then sources are found, relevant passages pulled out and the material put into a usable form. Shortening is one option, not a duty. Sometimes the full wording is both the simplest and the best.
Then the material gets used. That is where you see whether the preparation was enough. The system notices a gap, fetches another source or asks a question. Afterwards someone has to decide what should stay. A confirmed finding helps again later. A rejected assumption should stay recognisable as rejected. Confidential material needs its own rule.
Then come rules for access and use, and the handover to others. We also want to record what the preparation changed and which material led to a result. That gives a cycle: decide, find, extract, prepare, use, keep, govern, hand over, evidence and revisit. Those steps have to fit together.
Why look at them together? Because a change in one place shows up somewhere else. A very short summary is little help if an important condition is missing at the next step. Stored knowledge needs a route to correction. A shared package needs a clear recipient. The value appears when the work as a whole can continue.
Looking back later
LeanCTX also has context snapshots. These are saved states that can be tied to a project revision and signed. They help you look back at what material was there at the time. That is not the same as seeing a model think. And the same inputs do not automatically produce the same answer from an AI system later. [2]
Bundling material on purpose and passing it on is the second half of the idea. Context packages in LeanCTX exist for exactly that. Behind them sits a general design question. What does the next person or the next program really need? And what is better left where it is? [3]
Here technical precision meets human clarity. Good preparation names the goal. It makes the origin of important statements findable. And it separates the settled from the open. That helps a person check the work. It also helps a program carry on cleanly. The format can vary. The responsibility for the selection stays.
This is why we build LeanCTX. The decisions around context hang together and need a coherent implementation. For Thinkery it is the first answer to a larger belief: careful preparation should keep paying off. Anyone who later picks a different model or tool should be able to take that knowledge along and test it again. Good preparation may outlive a single answer.
From the Thinkery workshop
Our view of the work behind our products. Examples help explain the ideas. Links show which features already exist. Where we describe a goal, work is still ahead of us.
Edition of 9 September 2026