Developer A/B tests AI system prompt across 24 runs to sharpen local coding agent

·1 views

A developer running a local AI coding agent called FLASH, built on the Gemma 4 model, conducted 24-generation A/B tests to refine its system prompt for version Onyx 2.2. The update doubled the context window from 32,768 to 65,536 tokens after the developer measured actual usage costs rather than estimating them. Five new prompt sections were added covering how the model reads requests, names files, evaluates finished work, and suggests next steps. A key finding was that instructing the model to write files to disk without pasting full contents back into the reply significantly preserved the limited context window on local hardware. One new rule aimed at improving request interpretation had no measurable effect, highlighting that prompt engineering gains are uneven and require empirical testing.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

SSR Cache Isolation Flaw in Popular React Data Fetching Library Leaks User Data

A security researcher auditing a widely used React data fetching library discovered a cache isolation vulnerability affecting server-side rendering environments. The library initializes its cache as a module-level variable — a single shared Map object created once and reused across all incoming server requests in Node.js. Because modules are loaded once and shared globally, concurrent users hitting the server could inadvertently receive another user's cached data under the same cache key. The flaw requires no malicious input; two legitimate users making simultaneous requests is enough to trigger cross-user data exposure. The researcher found no built-in mechanism such as AsyncLocalStorage or per-request cache factories to enforce request-scoped isolation by default.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Developers Can Automate AI-Powered Upsells Inside Customer Support Chats

A technical guide published on DEV Community outlines how to build an automated upsell system that triggers personalized product offers during live customer support interactions. The system connects Intercom or Zendesk as the support front-end with OpenAI's language model to generate contextual product suggestions in real time. The workflow is orchestrated using n8n, an open-source automation tool, which routes conversation data to OpenAI, then fetches relevant products from a Shopify store and sends the customer a checkout link. Conversion events are subsequently logged into a CRM such as HubSpot or Salesforce, requiring no manual input from the support agent. Developers familiar with REST APIs are estimated to need six to eight hours to complete the full build.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How one dev team cut redundant CI runs in a merge queue to speed up delivery

A software development team found that their merge queue, adopted to speed up code delivery, was running a full test suite separately for each pull request waiting to land — meaning four queued PRs triggered four complete runs, even though the last already included the others. Their test suite comprised over 15,000 tests taking 13 to 19 minutes per run, with 93% of CI time consumed by the test step alone. The team discovered that the slowest 1% of tests — around 226 that spin up real processes and migrate databases — accounted for 41% of total test time, making simple test deletion ineffective. After moving compute to idle owned hardware, the same suite dropped from up to 16 minutes to 108 seconds. The article examines whether redundant runs in a merge queue can be safely eliminated and what approaches to stopping them are viable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Rebuilds AI-Generated Portfolio After Anonymous Feedback Called It 'Scary'

A developer spent nearly a week redesigning their portfolio site after anonymous user feedback and AI analysis revealed it looked entirely machine-generated. Using Google Gemini to audit the site, they found nearly every element — from color choices to layout — bore obvious AI tells. Over ten redesigns with Claude still produced results that felt like browsing archives rather than showcasing a person. The breakthrough came when the developer shifted the prompt from requesting a structured layout to asking for an 'artsy page,' which finally yielded designs they didn't dislike. They ultimately settled on a set of self-imposed design rules, including a single accent hue, contrasting fonts, and a focus on representing the human rather than cataloguing the work.

0 comments Read more at DEV Community