Code · Applied AI

Give AI a job, not a wish list

Evaluate a tool against a real task, including the time it takes to check and repair its output.

A new model is interesting. A repeatable improvement in your work is useful. Start the evaluation with one task you already do, and a clear definition of a good result.

Choose one bounded task.

Try structuring source notes, reviewing a draft, or classifying documents. Keep sensitive information out of a service until you understand its access and data controls.

Measure the whole workflow.

Include the time spent writing instructions, checking output, and correcting mistakes. A fast first draft that requires extensive repair may not save time.

Keep a human checkpoint.

Decide which claims need sources and which actions require your approval. Save a small set of examples so the same workflow can be tested again when tools change.

An illustrative workflow guide. No particular product, benchmark, or outcome is endorsed or guaranteed.

Subscribe

New essays, when they are ready.

Follow on Substack or add the RSS feed to your reader. Every essay is free to read.