Prompt Craftism

From the AI Research Compass wiki

Prompt Craftism: Carefully handcrafted prompts unlock most of a model's ability, so prompt writing is a skilled craft worth mastering word by word.

Beliefs

Prompt Craftism treats the prompt as an artisan’s workpiece. The right role, the right examples in the right order, the right phrasing of constraints and output format: each choice can swing results, and the craftsperson knows which ones from long practice. Where others see prompting as a temporary hack, the craftist sees a genuine skill, closer to editing or teaching than to programming, and keeps a well-worn notebook of templates that work.

It sits near the far Language edge, since everything happens in natural language, and a little below the middle, toward Harness, because the prompt is scaffolding around a fixed model rather than a change to the model itself. Its Orthodox shade marks loyalty to the mainstream large-model recipe: prompt craft was the defining practice of the early large-model era, and newer camps talk instead about context engineering, automatic prompt optimization or simply models that need less coaxing. The craftist is unbothered; handmade still sells.

Arguments

  • Few-shot examples, instructions and formatting choices can change task accuracy substantially without touching weights.
  • Prompting is cheap, immediate and accessible to non-specialists, so it is often the fastest path from a model to a working product.
  • A well-written prompt is human-readable documentation of intended behaviour, unlike opaque fine-tuned weights.

Critiques

  • Results are brittle: small wording changes or a new model version can break carefully tuned prompts.
  • Automatic prompt optimization and better instruction-tuned models erode the value of manual craft.
  • Critics argue prompt lore is often anecdotal and rarely evaluated with proper controls.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.71, y ≈ −0.81, paradigm ≈ −0.86.

Categories: Linguism · Paradigm: Orthodox

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data