14 prompts
Prompts that survive production
Not a list of clever phrasings. Each of these is written the way it would be written for a system that runs a million times a month — delimited, with the output contract stated and the edge cases decided.
And each one shows the version most people write, and the specific failure it produces. That part is the reason to read the page.
Writing
- Rewriting promptMaking something clearer without it coming back sounding like every other piece of text on the internet.prompt to rewrite text for clarity
- Product description promptProducing listing copy at volume where an invented feature is a returns problem and, in several jurisdictions, a legal one.prompt to write product descriptions
Analysis
- Document summary promptTurning a long report, transcript or article into something a person will actually read, without the summary quietly adding claims the source never made.prompt to summarize a document
- Meeting notes promptTurning an hour of transcript into the four things that actually have to happen, without inventing an owner for the task nobody claimed.prompt to summarize meeting notes into action items
- Sentiment analysis promptReading customer feedback at volume where "negative" tells you nothing and "negative about delivery, positive about the product" tells you what to fix.prompt for sentiment analysis
Code
- Code review promptGetting a second pair of eyes on a diff before a human reviewer spends theirs, where a list of style nitpicks is worse than no review at all.prompt for ai code review
- Test case generation promptFilling the gaps in a test suite, where the value is entirely in the cases the author did not think of.prompt to generate unit tests
- Code explanation promptLanding in a codebase you did not write and needing to understand a file well enough to change it safely.prompt to explain code
- Error debugging promptFacing a stack trace at the point where you have already tried the obvious thing and it did not work.prompt to debug an error message
Support
- Ticket classification promptRouting inbound messages to the right queue at volume, where the cost of a wrong label is a customer waiting in the wrong place for two days.prompt to classify support tickets
- Content moderation promptScreening user submissions at volume, where a wrong removal is a support ticket and a wrong approval is a much larger problem.prompt for content moderation
Data
- Data extraction promptPulling fields out of invoices, emails, CVs or support tickets into something a program can use, where a plausible wrong value is worse than an empty one.prompt to extract data from text into json
- Plain English to SQL promptLetting people ask questions of a database without writing SQL, where a query against a column that does not exist is better than a query that silently answers the wrong question.prompt to convert english to sql
- Data cleaning promptNormalising human-entered records — addresses, company names, dates, phone numbers — where a silent "correction" is indistinguishable from corruption.prompt to clean and normalize messy data
What makes these different
Most prompt libraries are a phrase and a promise. They work in the demo, and they fail in the ways nobody wrote down: the summary that rounds a figure, the extractor that invents an invoice number, the classifier that labels everything because it was never given permission to say “unclear”.
Every prompt here carries four things instead. The prompt itself, production shaped. The reason behind each rule, so you can adapt it rather than only paste it. The naive version and the exact failure it produces. And an honest list of what it still gets wrong.
The pattern underneath all of them
Read three or four and the same structure appears every time, because it is the structure that works: the task stated first, anything you did not write wrapped in delimiters, the output format shown rather than described, and every edge case decided in advance — what to do when the answer is not there, when two options fit, when the input is the wrong kind of thing entirely.
That last part is where most prompts fail. A model given a question its context cannot answer will answer anyway unless told what else to do. Almost every hallucination in a production system traces back to an edge case nobody decided.
Building one that is not here
The prompt generator assembles the same structure from an expert-authored blueprint for tasks these pages do not cover. Run the result through the prompt linter, which finds the unmeasurable words and the missing delimiters, and then the prompt review checklist before it reaches real traffic.
Before shipping any of them at volume, measure what they cost. A 400 token prompt at a million requests a month is a budget line, and the token counter and cost calculator turn that into a number in about a minute.