Prompt Templates — References

The evidence that surface form matters, the template languages themselves, and the standards that govern substituting values into text.

Why the exact rendered string matters

• Sclar, Choi, Tsvetkov and Suhr, 2023, on language models' sensitivity to spurious features in prompt design — the finding that semantically identical prompts differing only in separators, casing, and spacing produce substantially different measured accuracy. The direct argument for pinning whitespace behaviour and reviewing renderings rather than sources. Cited by name.
• Schulhoff and colleagues, 2024, The Prompt Report, a systematic survey of prompting techniques — useful as a catalogue of the structures a template family ends up needing to express. Cited by name.

The substitution step

• Unicode CLDR — the locale data behind number, date, and plural formatting; the reason a value rendered into prompt text can differ by region unless the formatting is pinned explicitly.
• ICU MessageFormat — the message syntax for plural and selection forms, and a worked example of separating a message's skeleton from its arguments. Cited by name; the documentation has moved between ICU releases.

Template languages

• Jinja documentation — in particular the undefined-value behaviour, autoescape semantics, and the whitespace-control options that decide whether a rendered prompt is stable across engine versions. Cited by name rather than linked; the documentation is versioned and its paths change.

No verified reference is recorded for prompt-specific templating practice — golden renderings, trust-typed holes, or fragment libraries for prompt text. The published work addresses whether a given wording works, not how a team maintains several hundred renderable prompt artefacts.