Until August, my work with Claude followed a hierarchy: one overarching session using the leading Claude model kept track of all my projects. Each project had an Opus-family architect and Claude Code for implementation. At the time, I barely trusted ChatGPT to help me phrase a paragraph.
Then, in two projects, I replaced both the architect and Claude Code with a single GPT-6 Astra session in Codex for each project. That worked well for a month. I then switched fully and moved the overarching session to Astra too. Its seemingly endless working context tempted me to let that one session handle every project down to the implementation details. That lasted two days.
My main lesson from these months: project knowledge, decisions and workflows need a lasting foundation of their own, so that different models and sessions can continue the work.
For me, that foundation is becoming increasingly local. I want to hold my own data, choose model versions and adopt updates once they have proved useful in my workflows. Local models can already handle bounded routine tasks. Frequent use can also make the economics attractive, with hardware and maintenance included in the calculation.
At the same time, the strongest cloud models still do more in my work: strategic reasoning, difficult architectural decisions and long, complex tasks. I expect that advantage to remain for a while. I want to use those capabilities deliberately and preserve the results in my own system.
The local landscape moves quickly too. A new model needs to fit my tasks, data and hardware. That is why I am developing a test environment for Folio. It helps me decide which changes actually improve the work. The latest synthetic comparison included Kolibri and Kev, an open decision model, alongside existing Gemma and Qwen profiles.
Connecting these two worlds is the central design question. One approach in Folio is the domain compass: local knowledge and open questions are assembled for the relevant strategy session. Its answer returns as a reviewable proposal and, after confirmation, can become part of the system's own memory. The aim is to carry context and decisions across models and sessions. I am still working on that connection.
The redesigned Aion Lumen website shows the current workshop: Folio starts with mail, invoices and appointments, with source references, memory (RAG using full-text and graphs) and an overview of recurring costs (AI-assisted transaction analysis and reconciliation) alongside them. The model comparison shows some of the testing behind this work.
Next, I want to test larger local models on the Mac Studio I have ordered. The repeated delivery delays and now higher price make me suspect that many others share my view of local AI's potential. Once it arrives, the next round of testing begins.
I now have Opus 5.5 review the work of OpenAI's Astra and Sol, with useful corrections and plenty of appreciation. That strengthens my confidence in the collaboration, much as the variety of my local models does. This is how I understand surfing the AI wave: finding stability through balance and clear roles for models from different providers. But the surfboard must belong to me and fit me.