Projects / 01 · Ongoing work
Building the
collaboration.
The model can draft, build and investigate. I’m working on the environment that makes its output easier to examine, question and improve.
- Focus
- AI harnesses
- Stage
- Building and iterating
- Perspective
- Human direction, inspectable work
The method
Build the system
around the model.
I build the context, tools, memory and handoffs that help a person understand the work before deciding what to do with it.
Judgment
Set the purpose.
Choose the next move.
Give it a frame.
- Context
- Tools
- Memory
- Handoffs
Do the work.
Draft. Build.
Explore possibilities.
Show what holds.
Sources. Checks.
Gaps still open.
The approach in practice
Understand the work.
Then make the call.
A useful review starts with what the result means, why it might be right, and what still needs checking. Those questions give a person something to base a decision on.
A result is ready.
What needs checking?
01 / CHECKRead the result. Check its reasoning and evidence.
02 / REVISEName what needs to change and why.
03 / APPROVEDecide whether it meets the purpose.
A visual example. These controls send nothing.
The work
Start by understanding the work.
I’m a chemical engineer. I start by trying to understand how a system works and what would count as an improvement. That is also how I approach AI: give it a specific piece of work, then examine what comes back.
My working environment includes a local inbox and project board, a harness that defines the model’s tasks and available tools, and ways to bring notes back into a session. I keep developing these tools as I use them.
Where it lives
On my own computer, where I can read it.
The inbox and the project board are files on my own computer. I can open them and see what a session was given and what it did. I think that matters because the direction of a project should stay with the person, even when the model or the tool changes.
The questions
What makes this useful in practice?
Context that survives.
How much can a new session pick up without making the person explain the whole project again?
Work you can examine.
Can a person trace the result back to the information it used, see what changed, and identify what still needs checking?
Room for human judgment.
What does the person need to understand to make this decision? What should happen when the answer is still unclear?
One consequence
Sometimes the product leaves AI out.
The Return Authorship app grew out of my experiences building with AI. It is deliberately a no-AI space for attention and choosing a direction.
The Return Authorship app →Follow the work
More of this belongs in public.
I write about what I build, where it goes wrong, and how my understanding changes.