Harness² vs Hand-Written Prompt Engineering
Last updated: October 19, 2026
When hand-written prompts are enough, and when a recursive harness earns its cost. A practical boundary between the two approaches.
Hand-written prompt engineering is not obsolete, and Harness² is not free. The useful question is where the boundary sits.
Where hand-writing wins
One-off tasks. When you will run something a handful of times, the loop of editing a prompt and re-reading the output is faster than setting up a harness, configuring a model and running a benchmark. Hand-writing also wins when you already know the shape of the answer and just need the model to format it.
Where the recursive approach wins
Repeated professional tasks. Every hand-edit is a single sample: you change one thing, look at one output, and decide. Harness²'s Propose–Probe–Compose step deliberately probes multiple candidates and compares behaviour across runs, which is what you would do by hand if you had the patience — and what nobody actually does at scale.
A rough test
| Ask yourself | If yes |
|---|---|
| Will this run once or twice? | Write the prompt by hand |
| Can I tell a good output from a bad one instantly? | Hand-writing is probably enough |
| Does this task repeat, with variation, many times? | A harness is worth the setup |
| Is "correct" expensive or slow to judge? | Behavioural comparison beats guessing |
| Am I editing the same prompt every week? | That loop is what Harness² automates |
The honest caveat. A harness adds moving parts: dependencies, model credentials, benchmark setup and run directories. The README's own demo exists precisely so you can see the loop before committing to that overhead. If you never get past step one of that test, stay with hand-written prompts.