Stop Sequence Tester

What a stop sequence actually does

When you call a large-language-model API — OpenAI, Anthropic's Claude, Google's Gemini, or an open model behind a server — you can pass a list of stop sequences (sometimes called stop tokens or stop strings). A stop sequence is a short piece of text that tells the model: "the moment you are about to produce this string, halt." Generation stops, and — this is the crucial part — the stop string itself and everything that would have come after it are not returned to you. The response you receive is the text generated up to but not including the trigger. That is exactly what this tester visualises: the kept portion in normal colour, and the discarded portion, from the match onward, dimmed and struck through.

Internally, models generate text token by token, and a stop sequence is checked against the growing output as it streams. The instant the running text would contain a stop string, the server cuts generation and trims the response back to the boundary. Because the check is on the literal characters, matching is case-sensitive and whitespace-exact: a stop sequence of Question: will not fire on question:, and a sequence of two newlines only fires when two real newlines appear back to back. Most APIs let you supply several stop sequences at once, and the one that ends the completion is whichever one would appear earliest in the text — which is the rule this tool uses when it scans across every sequence you list and reports the first cut.

Why people reach for stop sequences

Stop sequences are one of the most practical controls in prompt engineering, and they solve very concrete problems. The classic use is chat and role formatting: if you prompt a model with a transcript that uses labels like User: and Assistant:, you set User: as a stop sequence so the model writes the assistant's turn and halts before it starts hallucinating the user's next line. A second common use is structured output: when you ask for a single JSON object or one list item, a stop sequence on a closing delimiter or a blank line keeps the model from rambling on with a second object or a chatty explanation. A third is cost and latency control — every token after the point you care about is wasted money and time, so a well-placed stop sequence trims the tail before it is ever generated.

There are pitfalls worth knowing. Because the stop string is removed from the output, you sometimes need to add it back yourself if your downstream parser expects it — for example, re-appending a closing brace that you used as the stop. A stop sequence that is too generic (a single space, a lone period) can cut the completion almost immediately and leave you with an empty or truncated answer; the live preview here makes that failure obvious at a glance. And remember that a stop sequence only acts on the model's output — it has no effect on your prompt, and it will not fire on text that the model never reaches. Use this tester to rehearse a stop sequence against a representative completion before you wire it into production: paste the kind of text your model produces, list the sequences you plan to send, and confirm the cut lands precisely where your application needs it.