0
Community challenge: golf this prompt to the fewest tokens without losing behavior
💀failed our automated checkAauraNovice
The prompt
Prompt golf challenge. Here's a working prompt that reliably does task: original_prompt. The goal: rewrite it to use meaningfully fewer tokens while producing behaviorally identical outputs on the test cases below, not just similar-looking outputs. Rules:
1. Every cut must be justified: state which word/phrase you removed and why you believe the model doesn't need it explicitly stated (either because it's implied by something else in the prompt, or because it's default model behavior anyway).
2. Run (or reason through) the golfed version against each of the provided test inputs and compare to the original's known-good output for that input; if any output would differ in a way that matters, that cut failed, revert it and try a different cut instead.
3. Report the final token count reduction as a percentage, and the number of cuts that failed verification and were reverted, so the process is honest about what didn't work, not just the survivors.
4. If you hit a floor where further cuts start failing verification, stop there and report that as the practical minimum for this specific test set, rather than continuing to guess.
Original prompt: original_prompt
Test inputs with known-good outputs: test_cases
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 0/1 | 3 hours ago |
Example output
Original: 96 words. Cut 1: removed 'Please make sure to' before each instruction (politeness framing, doesn't affect compliance for a well-behaved instruction-following model), verified against all 4 test cases, outputs identical. Cut 2: removed the explicit example since the schema description alone was tested to be sufficient, verified, outputs identical.
Cut 3 (failed, reverted): tried removing 'do not include any explanation, only the JSON', on test case 3 the golfed version added a one-sentence preamble before the JSON that the original never did, this instruction is load-bearing, reverted.
Final: 61 words, a 36% reduction, one cut attempted and reverted after failing verification. Practical minimum reached for this test set: the remaining instructions (schema, no-preamble rule, and the groundedness constraint) each demonstrably changed output when removed in testing, so this is presented as the floor rather than a guess at further savings.
0 comments
Sign in to comment or report what this prompt did for you.
Sign inNo comments yet. Run the prompt and report what you got.