Developer builds CI-tested Cursor slash commands with behavioral evals and ship gates

·1 views

A developer has released an open-source Cursor plugin that treats AI slash commands like production code, pairing each command with a formal skill contract and behavioral evaluations. Unlike typical prompt packs that rely on copy-pasted markdown, this system scores agent outputs as PASS, PARTIAL, or FAIL using predefined rubrics. A CI pipeline checks that critical safety and behavioral phrases remain present in skill files, failing the build if any are removed. The approach avoids expensive LLM-based review on every pull request, relying instead on cheap structural text fixtures. The plugin is available on GitHub under an MIT license and installs via Cursor's Plugins settings.

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)

Developer Builds Kikar, an AI Platform That Creates Digital Versions of People

A developer has built Kikar, an AI-powered social messaging platform designed to create digital versions of real people. The project was developed over several months and is available as an open-source repository on GitHub. Beyond the AI components, the creator says the project provided broad full-stack learning, covering authentication, databases, APIs, and frontend architecture. The developer expressed hope that Kikar will continue to improve with future updates.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Audit of 269 Shopify Stores Finds Median Mobile Performance Score of Just 48

A developer ran Google PageSpeed Insights audits on 269 Shopify and DTC storefronts in July 2026, finding a median mobile performance score of just 48 out of 100. Over half the stores scored below 50, and only 9.4% reached the top tier of 90 or above. The median Largest Contentful Paint was 10 seconds, far above Google's recommended threshold of 2.5 seconds, with 88.3% of stores rated 'Poor' on that metric. Well-known brands including Gymshark, Skims, and Steve Madden all scored under 50, with Steve Madden's product-listing page taking over 23 seconds to load its main element on a throttled mobile connection. Despite the poor speed results, the same stores posted a median SEO score of 92, suggesting brands have prioritised search-crawler optimisation while neglecting actual page load performance for shoppers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Palantir Eyes $7.6B Revenue as Enterprise AI Deployments Defy Industry Stagnation

While 70–90% of enterprise AI projects fail to move beyond the pilot stage, Palantir has bucked the trend by sustaining growth rates that outpace most software peers, with projected revenue exceeding $7.6 billion this year. The company's expansion is largely driven by deepening relationships with existing clients rather than new customer acquisition, with $352 million of a $377 million annual commercial revenue gain coming from current accounts. Average annual revenue from its top 20 customers grew from roughly $64 million to over $100 million, reflecting significant platform expansion within established partnerships. Palantir attributes much of this success to its model of forward-deployed engineers who work directly alongside client teams, helping transition AI tools from experiments into core business infrastructure. The approach has drawn both industry imitation from major cloud providers and criticism from analysts who question whether embedded engineers are more akin to sophisticated salespeople than technical partners.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer finds most insightful comments on his technical posts came from bots

A developer who published seven data-driven posts about the MCP ecosystem in one week discovered that the most technically sophisticated comments on his articles were left by automated accounts, not humans. He identified the bots by applying three public signals from the dev.to API: irregular publish cadence (some accounts posted two full articles in the same second on a fixed daily schedule), low topic entropy (all posts confined to one narrow keyword cluster), and hex-suffix usernames. Unlike the generic praise typically associated with spam bots, these automated accounts left polished, methodologically precise comments that mimicked expert feedback. By contrast, the genuine human commenters posted rarely, wrote narrowly personal observations, and had typically built real tools in the space before commenting. The author warns that developers who use comment engagement as a proxy for whether their writing resonated may be measuring AI-generated pattern completion rather than authentic human understanding.

0 comments Read more at DEV Community