Blog posts tagged with aihttps://jamesst.one/https://jamesst.one/posts/the-other-ai-bubble-that-no-one-is-talking-about/https://jamesst.one/posts/the-other-ai-bubble-that-no-one-is-talking-about/import Detail from "../../components/Detail.astro"; Excuse my click baity title. I mean the AI Filter Bubble. <span>As the memory features of AI apps are getting <Detail><span slot="c">smarter</span><span slot="e">a larger context window</span></Detail>, I have recently found it harder to convice AI to not always recommend [nix](https://jamesst.one/tag/nix)! Ok that [example](https://chatgpt.com/share/69931fce-8be8-8002-bca2-dfeb7732be45) is specific to me. It has clearly remembered my love for it. But it got me thinking we are just making another [filter bubble](https://en.wikipedia.org/wiki/Filter_bubble). It's social media all over.</span> <br/> <span>Whilst _I_ don't think there is anything wrong with nix. Sometimes it doesn't hurt to hear other <Detail><span slot="c">minority</span><span slot="e">well at least less <a href="https://distrowatch.com/table.php?distribution=guixsd">popular</a></span></Detail> [lispy](https://guix.gnu.org/) opinions.</span>Mon, 16 Feb 2026 00:00:00 GMThttps://jamesst.one/posts/how-i-use-agents/https://jamesst.one/posts/how-i-use-agents/import Detail from "../../components/Detail.astro"; <div>Like many people, I have found <Detail><span slot="c">AI</span><span slot="e">LLM</span></Detail> agents to be increasingly useful.</div> Like people, some of the biggest power ups I have found are to give the agent a good environment. It works best when it: - has access to an **LSP** to allow it to quickly check type errors or just any error really. - can quickly run **lint** checks - has an existing **test** suite that it can incrementally add new tests for - in particular, I have found integration tests to be the most useful, that way you can let the agent go wild and not get bogged down by a particular implementation. It should not be scared to change - **Web dev tip**: Give it a browser it can use in the integration tests. - has a working **formatter** The beautiful thing is most agents auto run tests, code formatter and linters after every change. ----- { <div>My 2 cents above is based on my experience <Detail><span slot="c">"vibe coding"</span><span slot="e">"vibe coding", I also haven't truly been 100% vibe coding... I still look at the code it writes and then based on that either: <ul><li><b>completely undo</b> the change and ask it to have another go, maybe slightly tweaking my prompt</li><li>give it a follow up prompt.</li></ul> <b>though I think starting fresh seems to get better results.</b><br/>So I have been "vibe coding"</span></Detail> an app to help me keep track of my running niggles/ injuries. </div> } The agents really came alive when I made them write some integration tests, using an actual browser. eg: ```ts import { test, expect } from "@playwright/test"; test("homepage loads just added diagnosis", async ({ page }) => { await page.goto("/"); await expect(page).toHaveTitle("Current Diagnoses"); await expect(page.locator("diag 1")).toBeVisible(); }); ```Sun, 25 Jan 2026 00:00:00 GMT