<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Lazar Aćimović</title>
    <description>The latest articles on DEV Community by Lazar Aćimović (@lazar_aimovi_81401df3b8).</description>
    <link>https://dev.to/lazar_aimovi_81401df3b8</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4209005%2F50121e03-c7d0-404d-82eb-845cd27f009f.jpg</url>
      <title>DEV Community: Lazar Aćimović</title>
      <link>https://dev.to/lazar_aimovi_81401df3b8</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/lazar_aimovi_81401df3b8"/>
    <language>en</language>
    <item>
      <title>𝗚𝗼𝗼𝗱 𝗽𝗿𝗼𝗺𝗽𝘁 + 𝗯𝗮𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 = 𝗯𝗮𝗱 𝗼𝘂𝘁𝗽𝘂𝘁</title>
      <dc:creator>Lazar Aćimović</dc:creator>
      <pubDate>Sun, 01 Nov 2026 04:40:12 +0000</pubDate>
      <link>https://dev.to/lazar_aimovi_81401df3b8/--3i44</link>
      <guid>https://dev.to/lazar_aimovi_81401df3b8/--3i44</guid>
      <description>&lt;p&gt;Let's say you have an idea for an app. You've written a brutally good prompt to the AI, down to the smallest detail in terms of: &lt;br&gt;
What's the purpose of the app&lt;br&gt;
Defining the functionality.&lt;br&gt;
The user's role&lt;br&gt;
-restrictions&lt;br&gt;
-tech stack, etc. &lt;/p&gt;

&lt;p&gt;But you didn't give him enough 𝗰𝗼𝗻𝘁𝗲𝘅𝘁.  &lt;/p&gt;

&lt;p&gt;If the AI does not have information in front of it: &lt;br&gt;
-How the project is organized and what kind of architecture it uses.&lt;br&gt;
-What rules and conventions are already in place?&lt;br&gt;
-How the frontend, backend, and database are connected.&lt;br&gt;
-What are the rules of business that need to be followed?&lt;br&gt;
-What features have already been implemented and how they work.&lt;br&gt;
-which libraries to install and what to leave out&lt;br&gt;
-which rules to stick to&lt;/p&gt;

&lt;p&gt;It can easily happen that it generates an AI slop for you that later has to spend a lot of time to fix it. It can generate code to some extent that works half-as we expected, but that would be chaotic for further maintenance, especially for production.&lt;/p&gt;

&lt;p&gt;So, before an agent writes you the first line of code, make sure you have good context and a good prompt afterward.  &lt;/p&gt;

&lt;p&gt;Documents such as PRD.md, Tech.md, DB.md and files similar to them are a solution to this, as this extremely reduces the chance of bad output.&lt;/p&gt;

&lt;p&gt;Linkedin: &lt;a href="https://www.linkedin.com/in/lazar-a%C4%87imovi%C4%87-82a372301/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/lazar-a%C4%87imovi%C4%87-82a372301/&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  AI #SoftwareEngineering #AgenticEngineering #PromptEngineering #FullStackDevelopment #CleanCode #SoftwareDevelopment
&lt;/h1&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>softwaredevelopment</category>
      <category>softwareengineering</category>
    </item>
    <item>
      <title>𝗖𝗹𝗮𝘂𝗱𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗮 𝗣𝘆𝘁𝗵𝗼𝗻 𝘀𝗰𝗿𝗶𝗽𝘁 𝘁𝗵𝗮𝘁 𝗳𝗿𝗲𝗲𝗱 𝘂𝗽 𝟳𝟬 𝗚𝗕 𝗼𝗳 𝗱𝗶𝘀𝗸 𝘀𝗽𝗮𝗰𝗲 𝗳𝗼𝗿 𝗺𝗲</title>
      <dc:creator>Lazar Aćimović</dc:creator>
      <pubDate>Sat, 31 Oct 2026 10:35:57 +0000</pubDate>
      <link>https://dev.to/lazar_aimovi_81401df3b8/-4gfg</link>
      <guid>https://dev.to/lazar_aimovi_81401df3b8/-4gfg</guid>
      <description>&lt;p&gt;I've had a lot of trouble with my C drive lately, either because of Windows Update.&lt;/p&gt;

&lt;p&gt;As usual, I ran 𝗪𝗶𝗻𝗗𝗶𝗿𝗦𝘁𝗮𝘁 to scan the drive to see if I can delete anything other than .temp files. &lt;/p&gt;

&lt;p&gt;Seemingly, nothing, I thought. I need all of this data. Then, I was like: "Why would not I send scan results to claude"? He analyzed it and said that 70GB was only taken up by .dmp files and that it was safe to delete them.&lt;/p&gt;

&lt;p&gt;I've never heard of this kind of files until that moment. Anyway, I did a short research and realize that those are logs that the system prints whenever an application crashes, and the memory accumulated so much since I bought my current laptop just from those files.&lt;/p&gt;

&lt;p&gt;Then I told him to make me a python script that basically works by scanning the entire disk, finding files with the extension .𝗱𝗺𝗽, .𝗺𝗱𝗺𝗽, .𝗵𝗱𝗺𝗽, displaying the files to be deleted and deleting them only with subsequent confirmation (in case I need a log for possible analysis).&lt;/p&gt;

&lt;p&gt;The result - a revived C drive for 70GB. So yes, thanks to AI for postponing my purchase of a larger SSD. 😎&lt;/p&gt;

&lt;p&gt;Portfolio: &lt;a href="https://lazaracimovic.netlify.app/" rel="noopener noreferrer"&gt;https://lazaracimovic.netlify.app/&lt;/a&gt;&lt;br&gt;
GitHub: &lt;a href="https://github.com/LazarAcimovic" rel="noopener noreferrer"&gt;https://github.com/LazarAcimovic&lt;/a&gt;&lt;br&gt;
LinkedIn: &lt;a href="https://www.linkedin.com/in/lazar-acimovic-82a372301/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/lazar-acimovic-82a372301/&lt;/a&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  ClaudeCode #AI #Python #SoftwareDevelopment #DeveloperLife #Tech
&lt;/h1&gt;

</description>
      <category>ai</category>
      <category>automation</category>
      <category>python</category>
      <category>productivity</category>
    </item>
    <item>
      <title>How I Reduced Unnecessary AI Agent Context Usage by 50% (No More, No Less)</title>
      <dc:creator>Lazar Aćimović</dc:creator>
      <pubDate>Mon, 26 Oct 2026 14:55:31 +0000</pubDate>
      <link>https://dev.to/lazar_aimovi_81401df3b8/how-i-reduced-unnecessary-ai-agent-context-usage-by-50-no-more-no-less-2k32</link>
      <guid>https://dev.to/lazar_aimovi_81401df3b8/how-i-reduced-unnecessary-ai-agent-context-usage-by-50-no-more-no-less-2k32</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbtaya4ennd2lgq1kgk7l.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbtaya4ennd2lgq1kgk7l.png" alt=" " width="800" height="454"&gt;&lt;/a&gt;&lt;br&gt;
&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fur5qfle5qoih2xpqu4jg.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fur5qfle5qoih2xpqu4jg.png" alt=" " width="800" height="428"&gt;&lt;/a&gt;Let's get straight to the point - the before and after screenshots are at the end of the post.&lt;/p&gt;

&lt;p&gt;At first glance, people might think that 200k tokens of context is an impressive number. Not when simply reading documentation consumes over 60% of it before the agent even starts working on the task.&lt;/p&gt;

&lt;p&gt;A short backstory - I was refining my StudentPortal project by applying the principles of agentic engineering.&lt;/p&gt;

&lt;p&gt;Until now, it had 4 documents - PRD, Tech, UI, and DB - and the process worked by requiring every chat to read all 4 before implementing anything.&lt;/p&gt;

&lt;p&gt;As I kept adding more tasks and changes, a new agent couldn't even complete a subtask, let alone an entire task. On top of that, it would automatically trigger auto-compact, then mess everything up and forget what it had been working on. That's when I realized I had to change my approach.&lt;/p&gt;

&lt;p&gt;I solved this problem in &lt;strong&gt;2 steps&lt;/strong&gt;:&lt;/p&gt;

&lt;p&gt;I added an &lt;strong&gt;Overview.md&lt;/strong&gt; file, which every agent reads when starting (along with the AGENTS.md file). Essentially, it serves as a pointer map to specific documents. If a task involves business rules, the agent reads the PRD.md file. If a task is strictly related to the database, the agent reads the DB.md file. If it's related to architecture, it reads Tech.md. In other words, it never reads all 4 documents upfront.&lt;/p&gt;

&lt;p&gt;Using the &lt;strong&gt;grep tool&lt;/strong&gt; in combination with offset and limit. Now, you might be wondering: "Okay, the task involves creating a backend service. Does that mean the agent now has to read the entire tech.md file when it only needs the section related to backend architecture, not frontend architecture?"&lt;/p&gt;

&lt;p&gt;The answer is no - and that's exactly what grep is for. It solves this problem by finding the exact line number where the relevant section begins. Combined with offset and limit, the agent reads only that section instead of the entire file.&lt;/p&gt;

&lt;p&gt;The result - as you can see, after implementing this workflow, context usage after completing the first subtask was around 40% and a bit, which is practically half of what it was before.&lt;/p&gt;

&lt;p&gt;You can find the details of this system in the docs folder of my repository: &lt;a href="https://lnkd.in/dkumabfU" rel="noopener noreferrer"&gt;https://lnkd.in/dkumabfU&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If you have an app idea that is fairly complex, I'd be happy to hear that my system helped someone solve the same or a similar problem. Because, generally speaking, no matter how well you plan something, if the agent consumes too much context right from the start, building the application within that same context window becomes difficult.&lt;/p&gt;

&lt;h1&gt;
  
  
  AI #ClaudeCode #SoftwareEngineering #Documentation #ContextManagement #TechnicalWriting #Backend #WebDevelopment #DeveloperProductivity #AIAgents #PromptEngineering #AgenticEngineering
&lt;/h1&gt;

</description>
      <category>webdev</category>
      <category>ai</category>
      <category>dotnet</category>
      <category>react</category>
    </item>
  </channel>
</rss>
