This is the part nobody prices in. You tune prompts to a model’s quirks, the next version fixes those quirks, and your clever workflow quietly breaks. What’s held up for me is building on stable interfaces and treating prompts as replaceable, not sacred. Does the piece get into versioning?
Yes, prompts tuned to a model's quirks are the first thing to break when those quirks get fixed. Treating them as replaceable rather than sacred is the right instinct. This piece stays on the choosing side and doesn't get deep into versioning, but I wrote this one on treating prompts as software assets: versioning, evaluation, drift monitoring https://www.thedataletter.com/p/prompt-engineering
This is the part nobody prices in. You tune prompts to a model’s quirks, the next version fixes those quirks, and your clever workflow quietly breaks. What’s held up for me is building on stable interfaces and treating prompts as replaceable, not sacred. Does the piece get into versioning?
Yes, prompts tuned to a model's quirks are the first thing to break when those quirks get fixed. Treating them as replaceable rather than sacred is the right instinct. This piece stays on the choosing side and doesn't get deep into versioning, but I wrote this one on treating prompts as software assets: versioning, evaluation, drift monitoring https://www.thedataletter.com/p/prompt-engineering