R0021/2026-03-25/Q003/SRC02/E01

Research	R0021 — Prompt engineering definitions
Run	2026-03-25
Query	Q003
Source	SRC02
Evidence	SRC02-E01
Type	Factual

Anthropic's prompt engineering recommendations and measurability

URL: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

Extract

Anthropic's key recommendations:

1. Use XML tags to structure prompts — Structural. Specific format recommendation but no quantifiable success criterion.
2. Be clear and specific — Subjective. No metric for "clear."
3. Think step by step — Structural technique (chain-of-thought).
4. Use prompt chaining — Architectural pattern. Break complex tasks into subtasks.
5. Set roles via system prompts — Structural technique.
6. Handle large documents (20k+ tokens) with careful structure — Semi-quantifiable: specifies a token threshold.
7. Use few-shot examples — Structural but no guidance on how many.

Quantifiable recommendations found: 1 semi-quantifiable (20k+ token threshold for large document handling). Remaining recommendations are structural or subjective.

Anthropic also provides an interactive tutorial with 9 chapters — the tutorial teaches techniques but does not provide measurable success criteria.

Relevance to Hypotheses

Hypothesis	Relationship	Strength
H1	Contradicts	Almost no quantifiable recommendations
H2	Supports	Predominantly qualitative and structural guidance
H3	Supports	One semi-quantifiable threshold (20k tokens) amid qualitative guidance

Context

Anthropic's guidance is notably more structural than OpenAI's (XML tags, prompt chaining), but equally non-quantifiable in terms of measurable outcomes.