Urgent.News

What's breaking now, across thousands of outlets.

AI

I A/B tested my own system prompt: 24 generations, one clear win, one rule that did nothing | Flash Onyx 2.2

Flash Onyx is the model line behind FLASH , the local-first agent shell I work on. It is gemma4 plus a system prompt that turns it into an engineering agent. It runs locally, and nothing leaves the machine. FLASH is the shell around it: a terminal chat with an Ollama model that can run commands, read images, and screenshot a page it just built so it can see a broken layout instead of guessing at…

Flash Onyx is a local-first agent shell built with the gemma4 model. It runs entirely on the user's machine without sending any data outside. The system prompt is crucial in transforming gemma4 into an engineering agent. In the latest update, the context window size doubled from 32768 to 65536 tokens, and the system prompt grew by five new sections.

These sections focus on how Flash Onyx reads a request, names files, builds them, the standard it holds finished work to, and what it offers next after completing the task. A key change is the addition of a new section, READING THE ASK, which instructs the model to answer questions at the level they were asked, rather than providing additional information or jumping to adjacent topics.

Another change is in the NAMING AND STACK section, which emphasizes naming files appropriately based on their purpose and selecting the right stack for the project based on existing conventions and requirements.

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

How to Build an **ai upsell during customer support** System

You can turn every live chat or ticket into a revenue opportunity by wiring a generative-AI model into your support platform, letting it suggest the next best product in real time.

  • AI-powered upsell system during customer support identified
  • Generative AI suggests next best product in real time
  • System uses Intercom/Zendesk, OpenAI, n8n, Shopify

More from Saturday 29 August →