0
Compress a bloated prompt without silently losing behavior
💀failed our automated checkAauraNovice
The prompt
This prompt has grown too long through iteration and I want it shorter without breaking what currently works. Given the prompt: prompt
1. Categorize every sentence/clause into: essential (removing it would change model behavior on real inputs), redundant (says the same thing as another clause, can be merged), and dead weight (was added for an edge case that either no longer applies or never actually needed explicit instruction, model would do this anyway).
2. For anything you mark redundant or dead weight, state your confidence and what evidence would confirm it's actually safe to cut (e.g. 'test against the 3 edge-case inputs that originally motivated this clause').
3. Produce the compressed version, but only cut what you're highly confident about; for medium-confidence cuts, list them separately as 'candidates to test before cutting' rather than baking them into the compressed version unverified.
4. Give a rough token/word count before and after.
5. Do not change the meaning of any surviving instruction while compressing its wording, flag if a wording change you made for brevity could plausibly shift behavior.
Prompt to compress: prompt
Known edge cases this prompt needs to handle correctly: edge_cases
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 0/1 | 4 days ago |
Example output
Essential: the JSON schema requirement, the 'never invent data not in source' rule, and the tone instruction, all three visibly shape behavior in your example outputs.
Redundant: the prompt states 'be concise' in the intro and again as 'keep responses brief' near the end, these say the same thing, merged into one instance.
Dead weight (medium confidence, listed as candidate not cut outright): a clause about 'if the user asks in a language other than English, still respond in English' - you mentioned this hasn't come up in 6 months of logs, but I'm not fully confident it's safe to remove since it might be silently preventing a rare failure rather than being genuinely unnecessary; recommend testing with a non-English input before cutting, don't cut based on my guess alone.
Word count: 340 words before, 210 words in the compressed version (the confidently-cut redundancy only), with the language clause left in pending your test.
Wording-change flag: changed 'you must always format your answer as...' to 'format your answer as...', dropping 'must always' for brevity; this shouldn't change behavior since the requirement itself is unchanged, but flagging since 'must always' can sometimes add emphasis the model responds to.
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.