Urgent.News

What's breaking now, across thousands of outlets.

AI

The Prompt Changed and There Is No Commit for It

The answers got worse on Tuesday. Nothing shipped on Monday. The code is identical to the code that worked last week. The branch has not moved. Somebody edited the prompt in a browser console on Friday afternoon. There is no commit. No author. No diff. Nothing to revert to. The sentence that decides what your product says to customers is the only part of your system with no history. It got there…

On Tuesday, the quality of generated answers deteriorated noticeably. No release shipped on Monday. The underlying code remained unchanged since the preceding week. The branch had not progressed. A prompt alteration was made on a browser console on Friday afternoon, yet there was no record of a commit, author or difference between versions.

The sentence that determines how your product communicates with customers lacks any historical documentation. It emerged as an experiment originally. Once copied into the feature, it was treated as code without any formal treatment. A deleted clause that lengthened the response was responsible for price quoting restrictions. A truncated example aimed at three downstream systems was eliminated for brevity.

The use of the word 'concise' reduced the length of every response and removed a disclaimer. The model behind the prompt also changed due to the request for the latest version, which is inherently dynamic. As such, it is essential to treat the prompt as a file in the repository, shipped, reviewed and rolled back with releases. The model identifier must be pinned, documented as a specific version, and modified deliberately on a monitored day.

Maintain thirty real inputs and the constraints the answers must satisfy, not performance scores. The system should refuse to quote prices, return structured output and abstain from promising refunds. Before deployment, these properties must be tested. Furthermore, record the prompt version and the model that produced each response to enable accurate identification of the prompt and model responsible when a customer presents a screenshot.

Never allow a pricing rule to be altered through an unrecorded text box. In the current situation, that is precisely what is happening.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

I built guardrails for Claude Code after it told me a fix "worked" when it didn't

Just shipped this after getting burned one too many times: Claude Code Guardrails Pack — free hooks and skills that stop an AI coding agent from claiming work is done when it never actually checked.

  • Guardrails prevent Claude Code from falsely claiming code is fixed without verification.
  • Includes unlazy skill, session state snapshot, and safegitguard.py script.
  • Pre-commit-secret-scan.sh hook blocks commits with sensitive API keys.

More from Thursday 17 September →