Writing · Published
Notes on Useful AI
A working set of principles for using language models in serious knowledge work without surrendering judgment.
I am less interested in whether an AI system appears intelligent than in whether a workflow produces better, more inspectable work.
A few provisional principles
- Give the model artifacts, not just instructions.
- Keep consequential decisions visible and reversible.
- Separate collection, interpretation, and synthesis.
- Verify outputs in proportion to the cost of being wrong.
- Prefer workflows that improve human understanding as they run.
These principles are not a policy. They are a practical filter for experiments: does this arrangement help someone see and decide, or does it merely create more fluent output?