https://sumnerevans.com/tags/llms/Recent content in LLMs on Sumner EvansHugoen-usSat, 26 Sep 2026 13:08:43 -0600https://sumnerevans.com/posts/software-engineering/sgil/Fri, 25 Sep 2026 06:25:00 -0600https://sumnerevans.com/posts/software-engineering/sgil/<div class="admonition abstract"> <div class="admonition-header"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 384 512"><path d="M64 0C28.7 0 0 28.7 0 64L0 448c0 35.3 28.7 64 64 64l256 0c35.3 0 64-28.7 64-64l0-288-128 0c-17.7 0-32-14.3-32-32L224 0 64 0zM256 0l0 128 128 0L256 0zM112 256l160 0c8.8 0 16 7.2 16 16s-7.2 16-16 16l-160 0c-8.8 0-16-7.2-16-16s7.2-16 16-16zm0 64l160 0c8.8 0 16 7.2 16 16s-7.2 16-16 16l-160 0c-8.8 0-16-7.2-16-16s7.2-16 16-16zm0 64l160 0c8.8 0 16 7.2 16 16s-7.2 16-16 16l-160 0c-8.8 0-16-7.2-16-16s7.2-16 16-16z"/></svg> <span>TL;DR</span> </div> <div class="admonition-content"> <p>The chat-based workflow encouraged by existing AI harnesses results in a fragmented source of truth for the desired functionality of the software.</p> <p>To centralise the source of truth, we should use Markdown as Source (MaS). We should adopt a Specify, Generate, Inspect, Loop (SGIL) workflow which puts human-written Markdown files stored in source control front-and-centre.</p> </div> </div><h2 id="my-current-ai-workflow"> My Current AI Workflow <a class="permalink" href="#my-current-ai-workflow" aria-label="Permalink to My Current AI Workflow" ><!-- tags: [zoom-filled, link, url, connection, hyperlink, reference, attach, join, typography, writing] category: Text version: "1.0" unicode: "eade" --> <svg aria-hidden="true" class="hi-svg-inline" fill="none" height="1em" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"> <path d="M9 15l6 -6" /> <path d="M11 6l.463 -.536a5 5 0 0 1 7.071 7.072l-.534 .464" /> <path d="M13 18l-.397 .534a5.068 5.068 0 0 1 -7.127 0a4.972 4.972 0 0 1 0 -7.071l.524 -.463" /> </svg> </a> </h2> <p>I <em>hate</em> my AI workflow. I don&rsquo;t know how everyone else is using AI, but currently my workflow looks something like:</p> <ol> <li>Open Claude and type a chat message indicating what I want it to do.</li> <li>Look at the code or plans it generates and test the functionality of the code it produced.</li> <li>Type a chat message clarifying my ask or correcting the AI&rsquo;s assumptions. Normally this involves telling it what I dislike about the code it generated.</li> <li>Go to 2.</li> </ol>