R0021/2026-03-25/Q003/SRC03/E01

Research	R0021 — Prompt engineering definitions
Run	2026-03-25
Query	Q003
Source	SRC03
Evidence	SRC03-E01
Type	Factual

Google's prompt engineering recommendations and measurability

URL: https://ai.google.dev/gemini-api/docs/prompting-strategies

Extract

Google's key recommendations:

1. Be precise and direct — Subjective. But Google notes "most successful prompts tend to average around 21 words" — a rare quantifiable data point.
2. Use PTCF framework (Persona, Task, Context, Format) — Structural framework.
3. Always include few-shot examples — Structural. Strong recommendation ("always") but no quantity specified.
4. Use consistent delimiters (XML or Markdown) — Structural.
5. Keep temperature at 1.0 for Gemini 3 — Quantifiable. Specific parameter value with rationale.
6. Use responseSchema for structured output — Structural/technical.
7. Iterate: start simple, measure, adjust — Process recommendation.

Quantifiable recommendations found: - Temperature = 1.0 (specific parameter) - Average prompt length ~21 words (data point, though described as a tendency) - "Up to 99% retrieval accuracy on structured data, ~95% on mixed-content PDFs" (performance claim)

Google provides the most quantifiable guidance of the four vendors (3 measurable data points), but the majority remains qualitative.

Relevance to Hypotheses

Hypothesis	Relationship	Strength
H1	Contradicts	Even the most quantifiable vendor has <30% measurable recommendations
H2	Supports	Majority of recommendations are still qualitative
H3	Supports	Google demonstrates that some quantification is possible, making the absence from other vendors more notable

Context

Google's temperature=1.0 recommendation is notable as perhaps the only truly engineering-grade specification across all four vendors: it specifies an exact parameter value with an explanation of why deviation causes problems.