Smart Quote to Straight Quote Sanitization Prompt

This prompt is for developers and platform engineers whose code parsers, SQL engines, configuration loaders, or CLI tools choke on curly quotes (“ ” ‘ ’) that appear in model-generated text. The job-to-be-done is a deterministic character-level sanitization: convert all smart quotes and similar typographic characters into their ASCII straight-quote equivalents (" and ') without altering any other content. The ideal user is an integration engineer or backend developer who is piping model output into a system with a strict ASCII or Latin-1 contract and cannot afford parser failures, syntax errors, or silent data corruption caused by non-standard quote characters.

Use this prompt when the model output is structurally valid but contains typographic quotes that break a downstream parser. This is common when models generate JSON values, SQL string literals, shell commands, CSV fields, or code snippets where curly quotes are interpreted as syntax errors. The prompt is designed for post-generation repair, not for preventing the model from producing smart quotes in the first place. It assumes you have already received the output and need a reliable normalization step before the text enters a system that rejects non-ASCII quote characters. Do not use this prompt when the text requires locale-aware apostrophe handling (e.g., French or Italian elision where a curly apostrophe carries semantic meaning that should be preserved) without first adding explicit rules for those cases.

Before wiring this prompt into a production pipeline, define your eval criteria clearly. A successful sanitization means every ‘ and ’ becomes ', every “ and ” becomes ", and no other characters are modified. Common failure modes include false-positive corrections where a curly quote inside a code comment or a string literal that intentionally contains a typographic quote is flattened, or false negatives where uncommon Unicode quote variants like ‹› or «» are left untouched. If your downstream system has additional constraints—such as rejecting all non-ASCII characters—pair this prompt with an ASCII-only fallback prompt. For high-risk pipelines where quote corruption could cause SQL injection or shell command alteration, always run a diff check between the original and sanitized output and log any changes before the sanitized text is used.

This prompt template is designed to be dropped directly into your post-processing pipeline. It instructs the model to act as a precise text sanitizer, converting only typographic quotation marks and apostrophes into their ASCII equivalents while preserving all other content, including intentional formatting, code blocks, and locale-specific punctuation. The template uses square-bracket placeholders for the input text and any optional constraints you need to pass at runtime.

textCopy
Sanitize the following text by converting all typographic (curly/smart) quotation marks and apostrophes to standard ASCII straight quotes. Follow these rules exactly:

1. Convert left double quotes (“) to straight double quotes (")
2. Convert right double quotes (”) to straight double quotes (")
3. Convert left single quotes (‘) to straight single quotes (')
4. Convert right single quotes (’) to straight single quotes (')
5. Do NOT modify any other characters, including:
   - Standard straight quotes and apostrophes already present
   - Code blocks, inline code, or any text inside backticks
   - Intentional typographic characters like primes (′ ″) or angle quotes (« »)
   - Locale-specific punctuation such as German low quotes („ “)
   - Any non-quote characters, whitespace, or line endings
6. Return ONLY the sanitized text with no additional commentary, explanation, or wrapping

[INPUT]

[CONSTRAINTS]

To adapt this prompt, replace [INPUT] with the raw text from your model's output. Use [CONSTRAINTS] to pass any additional rules at runtime, such as preserving specific quote styles for certain locales or excluding particular sections from conversion. If you need to handle locale-aware apostrophes (e.g., French elision, Italian apostrophes), add those as explicit exceptions in the constraints block. For high-risk contexts like SQL query generation or code that will be executed, always validate the output through a secondary ASCII-only character filter before ingestion.

The smart quote sanitization prompt is designed to be a post-processing repair step in a pipeline, not a standalone user-facing feature. It should sit immediately after the primary model call and before any parser, SQL engine, config loader, or serialization step that is known to choke on curly quotes. The prompt expects a single string input ([INPUT]) and returns a sanitized string. Do not use this prompt on user-facing display text where typographic quotes are intentional and desirable—it is strictly for machine-to-machine boundaries.

Integration pattern: Wrap the prompt call in a lightweight function that accepts the raw model output string and returns the sanitized string. The function should: (1) Check if the input contains any curly quote characters (\u2018, \u2019, \u201C, \u201D) before calling the LLM—if none are present, return the input immediately to avoid unnecessary latency and cost. (2) Call the sanitization prompt with the input. (3) Validate the output by confirming no curly quotes remain. (4) If curly quotes persist, retry once with an explicit error message appended to the prompt context. (5) If the retry also fails, log the failure with the original and sanitized strings, and fall back to a deterministic regex replacement (text.replace(/[\u2018\u2019]/g, "'" ).replace(/[\u201C\u201D]/g, '"')) as a safety net. The LLM path handles locale-aware apostrophe disambiguation (e.g., preserving ' in contractions vs. replacing opening/closing single quotes); the regex fallback is a blunt instrument that trades nuance for reliability.

Model choice and latency: This is a low-complexity string transformation task. Use the smallest, fastest, cheapest model available in your stack that can follow the instruction reliably—typically a lightweight model like GPT-3.5 Turbo, Claude Haiku, or an open-weight 7B-class model. Do not route this to a large reasoning model. Set a short timeout (2-3 seconds) and treat timeouts as a signal to fall back to the deterministic regex path. Logging and observability: Log every invocation with: input length, whether curly quotes were detected, whether the LLM path or regex fallback was used, latency, and a boolean sanitization_success flag. This lets you monitor how often the LLM path is actually needed and whether the regex fallback is carrying too much load. Testing: Before deploying, run a golden test set of 50+ strings covering: mixed curly and straight quotes, nested quotes, locale-specific apostrophes (French l'école, Italian dell'arte), empty strings, strings with no quotes, and strings with only straight quotes. Assert that the output contains zero curly quotes and that legitimate apostrophes in contractions are preserved as straight single quotes, not removed or replaced with spaces.