Where the engineering belongs
A useful AI learning journey should change layers when the responsibility changes — not simply when the implementation becomes more technical.
Why I separated Prompt, Context, Harness, Loop and Graph Engineering across Work With AI, Give AI Work and Run AI.
Nothing. This is a record of a curriculum and mental-model correction.
Follow the reasoning, then use the linked stable guides for the current distilled version.
A clearer boundary between collaborating with AI, delegating responsibility to AI and operating a system of AI responsibilities.
Designed / curriculum decision. This note records a learning-architecture correction. It is not a production benchmark or a claim that one implementation pattern has been experimentally proven superior.
What I initially grouped together
When I started expanding the technical side of the learning library, the natural builder progression looked like:
Prompt
↓
Harness
↓
Loop
↓
GraphThat progression is useful technically, but it created a curriculum problem. It made the engineering terms look like a separate fourth course and blurred where Prompt Engineering and Context Engineering actually belong.
The correction
| Programme | Engineering depth | Responsibility shift |
|---|---|---|
| Work With AI | Prompt Engineering | I am still doing the work with AI. I need to express intent and steer deliberately. |
| Work With AI | Context Engineering | I am still doing the work with AI. I need to decide what reality and evidence it should work from. |
| Give AI Work | Harness Engineering | I am handing over a responsibility. The agent now needs a repeatable working environment. |
| Give AI Work | Loop Engineering | I want the responsibility to continue without me repeatedly deciding every next step. |
| Run AI | Graph Engineering | I am operating several responsibilities whose interactions, authority and outcomes must be owned as one system. |
The course boundary should move when responsibility moves. Technical sophistication alone is not a good curriculum boundary.
Why Context Engineering moved out of Harness
Harness Engineering still has to deliver context. A coding/task agent needs to find project rules, current work, source precedence and task-specific material reliably.
But that is an operational question:
How does the delegated agent reliably get the context it needs every run?
The more fundamental Context Engineering question appears earlier:
What context should the model have for this work, which source should be trusted, and what should remain unknown?
That belongs in Work With AI. The Harness consumes that discipline and makes it repeatable.
What this changed on 9order
- Work With AI now owns Prompt Engineering and Context Engineering.
- Give AI Work owns Harness Engineering and Loop Engineering.
- Run AI owns Graph Engineering.
- AI Field Guide is now the cross-course index and journey record, not another programme.
What I want to preserve as the journey continues
The stable guides should change when I learn something better. But a field note like this should preserve the decision and evidence that existed at this point in the journey.
Future build experiments can challenge this model. If that happens, the right response is not to quietly rewrite this note. The newer field note should explain what changed, while the stable field guide is updated to the current view.
Prompt and Context shape collaboration. Harness and Loop make delegation dependable and autonomous. Graph makes a system of responsibilities operable.