Use four buckets: Now, Next, Later and Ignore for now. Place a skill according to the work it enables, then revise the placement when your responsibilities change. “Ignore for now” preserves attention without making a permanent judgement about the technology.
| Current direction | Now | Next | Later | Ignore for now |
|---|---|---|---|---|
| Business or operations | Safe use, task framing, evidence checks | Repeatable templates and approved automation | APIs for repetitive work | Model training and CUDA |
| Software to AI applications | Model calls, schemas, evaluation baseline | Authorized retrieval and narrow tools | Agents or MCP when needed | Foundation-model pretraining |
| Infrastructure to AI platforms | Inference, memory, containers | Serving, monitoring and load testing | Distributed inference and profiling | Unrelated prompt-framework churn |
For each work cycle, name the problem, the current baseline and the improvement you expect to observe. “Learn agents” is too broad. “Reduce the time to produce a reviewed incident summary while preserving all material evidence” gives you something you can actually test.
Start with ten representative examples you can judge. Record the current effort and the errors that matter. Try the simplest approved design, inspect what fails and change only what the evidence points to. Keep a short decision record so an approach you rejected for a reason does not return later under a new product name.
When a new technology appears, first ask what job it performs. Then ask what it replaces, which constraints it changes, and what migration or operating work it introduces. Your evaluation should be able to detect the benefit being claimed. A promising demonstration is a reason to test, not evidence that your system needs the technology.
There is another useful question: what happens if you wait six months? If very little changes, your attention may be better spent finishing the current workflow. If a new requirement makes the technology necessary, move it into Now and give it a concrete test.
The five labs in chapters 27 and 28 are designed for that kind of test. Completing one carefully, including the failed cases, is more informative than following several tutorials whose outputs you never check.