<?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: Mugi</title>
    <description>The latest articles on DEV Community by Mugi (@mugitaiyaki).</description>
    <link>https://dev.to/mugitaiyaki</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%2F4254244%2Fc2be8443-683a-47bd-9d6f-1acb607e8106.png</url>
      <title>DEV Community: Mugi</title>
      <link>https://dev.to/mugitaiyaki</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/mugitaiyaki"/>
    <language>en</language>
    <item>
      <title>More Than Prompts: Context and Harness Engineering</title>
      <dc:creator>Mugi</dc:creator>
      <pubDate>Fri, 20 Nov 2026 13:07:01 +0000</pubDate>
      <link>https://dev.to/mugitaiyaki/more-than-prompts-context-and-harness-engineering-3lo4</link>
      <guid>https://dev.to/mugitaiyaki/more-than-prompts-context-and-harness-engineering-3lo4</guid>
      <description>&lt;p&gt;AI never tells you why it stopped.&lt;/p&gt;

&lt;p&gt;While I was building a game in Unity with Claude, work would sometimes stop with no report at all. The reason was simple: it was waiting for my go-ahead. No amount of prompt editing fixed that kind of stall.&lt;/p&gt;

&lt;p&gt;When I started delegating development to AI, I spent a while polishing how I wrote instructions. It helped. But the two things that changed my results were different: context engineering and harness engineering. I learned the names later. This post organizes what I was already doing.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Context engineering: designing the information you hand over
&lt;/h2&gt;

&lt;p&gt;An AI's output is bounded by the information it receives. What it doesn't know, it fills in with guesses. Outdated assumptions stay in force. So before any single instruction, what matters is how you arrange the whole body of information the AI works from.&lt;/p&gt;

&lt;p&gt;There is research behind this. Anthropic, the company behind Claude, defined context engineering in a September 2025 engineering post as curating and maintaining the optimal set of information for the AI to work with, including everything outside the prompt itself. The premise is that context is a finite resource. As conversations and documents pile up, the model's ability to recall information accurately degrades a little at a time. The phenomenon is called context rot.&lt;/p&gt;

&lt;p&gt;Here is what I did, in three parts.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;I keep a change log and have each new AI session read it. It records what changed, when, and why, so handovers need almost no explanation.&lt;/li&gt;
&lt;li&gt;Rules that must always hold go into memory: "Reply in Japanese." "Take a backup before changing anything." Promises that should survive across conversations.&lt;/li&gt;
&lt;li&gt;I keep the documents I hand over short. Long documents burn tokens fast.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The first item matches what Anthropic's post recommends as structured note-taking: have the AI write notes outside the conversation and read them back when needed. My change log is exactly that. I was doing it before I knew the name.&lt;/p&gt;

&lt;p&gt;There is also a format to how I give instructions. I don't throw raw thoughts at the AI. I hand over four things as a set: what I want to do now, what I'm stuck on, the current state, and the scope of the change. It sounds obvious, but skipping the obvious is what creates extra rounds of back-and-forth. A practical guide I found online organizes instructions as goal, current state, conditions, and a definition of done. Lining that up against my four, the piece I was missing was the definition of done. Saying up front what "finished" looks like. I'm adding that.&lt;/p&gt;

&lt;p&gt;The longer a conversation runs, the more the overall arrangement of information matters, compared to any single instruction.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Harness engineering: designing the system it works in
&lt;/h2&gt;

&lt;p&gt;A harness is the equipment that lets a horse work. In AI terms, it means the framework around the model: tools, files, permissions, checks, memory. LangChain defines an agent as Model + Harness. Everything that is not the model is the harness.&lt;/p&gt;

&lt;p&gt;There are numbers behind this too. LangChain improved their harness while keeping the model fixed, and their score on a coding benchmark (Terminal-Bench 2.0) rose from 52.8 to 66.5. The environment alone, with no model change, moved the result that much.&lt;/p&gt;

&lt;p&gt;What I set up looks like this.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A path for the AI to operate the Unity editor: compile, read errors, play, capture the screen.&lt;/li&gt;
&lt;li&gt;A way to build scenes from code, so running the same steps produces the same result every time.&lt;/li&gt;
&lt;li&gt;Backups before changes, and a flow where a different AI reviews the code after completion.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key point: the range the AI can verify by itself becomes the range you can safely delegate. An AI that writes code and stops is a completely different workload from one that compiles, reads the errors, and fixes them. Anthropic's Claude Code best practices say the same thing: giving the AI a way to verify its own work is the single highest-leverage move.&lt;/p&gt;

&lt;p&gt;Day to day, the core habit is fixing the environment after every failure. Mitchell Hashimoto, the creator of Ghostty, wrote that whenever an agent makes a mistake, you engineer the environment so the mistake can't happen again. My own rules about batch mode and about never showing confirmation dialogs were added the same way, after something went wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. What this looks like in Unity
&lt;/h2&gt;

&lt;p&gt;In Unity, both sides take concrete shape.&lt;/p&gt;

&lt;p&gt;On the context side, I learned it through Unity AI. Feed it Japanese text and it spends tokens wrestling with the language on its own. Long specification documents were the worst. Enough to make me want to cheat on it with Claude. Splitting documents into short pieces helped both cost and accuracy.&lt;/p&gt;

&lt;p&gt;On the harness side, it's the editor path and building scenes from code. Unity scenes are designed to be edited by hand on screen, so an AI needs machinery to touch them. Only once the AI can run the full loop itself (compile, check errors, play, capture the screen) can you delegate everything from writing to verifying.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. They nest
&lt;/h2&gt;

&lt;p&gt;The three terms nest inside each other. A prompt is part of the context. The context flows inside the harness.&lt;/p&gt;

&lt;p&gt;Instructions, then the whole body of information, then the whole working system. When a better prompt stops helping, shift your attention one level outward.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;Claude itself is excellent, and we get along well. That's exactly why the remaining difference shows up in how you hand over information and how you design the system around it.&lt;/p&gt;

&lt;p&gt;Models will keep getting smarter. As they do, the gap moves to the design of the information and the working environment.&lt;/p&gt;

&lt;p&gt;The hands-on record of this work is in my earlier post about letting an AI agent drive the Unity editor to build a VR app. That one is the example; this one is the reasoning.&lt;/p&gt;

&lt;h2&gt;
  
  
  More from this series
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The example: I Let an AI Agent Drive the Unity Editor to Build a VR App
&lt;a href="https://mugitaiyaki.substack.com/p/i-let-an-ai-agent-drive-the-unity" rel="noopener noreferrer"&gt;https://mugitaiyaki.substack.com/p/i-let-an-ai-agent-drive-the-unity&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The workflow: Building a Unity Game with Claude: Why Workflow Beat Prompting
&lt;a href="https://mugitaiyaki.substack.com/p/building-a-unity-game-with-claude" rel="noopener noreferrer"&gt;https://mugitaiyaki.substack.com/p/building-a-unity-game-with-claude&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>unity3d</category>
      <category>ai</category>
      <category>gamedev</category>
      <category>claude</category>
    </item>
  </channel>
</rss>
