←Ben’s Bites Waybackbensbites.com/p/what-makes-a-good-agentsmd19 Mar 2026sent to 157,812opened 32.5%clicks 10,072unsubs 146← 17 Mar · 21 Mar → · random

AGENTS.md / CLAUDE.md are instruction files that get pre-loaded before your conversation even starts.

Every model has a system prompt - instructions from the model/product makers. Your .md files get added to this. The model now has its system prompt + your instructions pre-loaded.

CLAUDE.md is specific to Claude products. AGENTS.md works across Codex/Droid/Pi/most other tools. You can ‘symlink’ these - ie link my AGENTS.md to CLAUDE.md so that I only ever deal with one set of instructions and any agent will pick it up correctly. (Just ask your agent to symlink them)

What goes into your AGENTS.md?

We previously thought including your tech stack, key files, etc., as like a mini-map for your agent was the right approach. That’s what agents add if they create it.

But there was a study that showed it hurt performance and increased cost by 20% (using extra tokens). The agent can figure out the tech stack, key files, commands, and architecture very easily and quickly.

Instead, it should be pretty empty. It should be your preferences and nudges to correct agent behaviour.

  - When building, open a browser with agent-browser skill and test before sending me a URL (to catch bugs)
  - Use the Exa web search tool for web search
  - Always write planning files in ~/[project-name]/plan/
  - I can't code, so explain things in simple terms
  - Record a video of your output so I can see exactly what you tested

I picked up a tip from Dex222, who says wrapping sections in conditional blocks is helpful:

<important if="simple web page">
  - No spec needed
  - Create 3 designs before choosing one
  - Must have dark/light mode switcher
</important>

I often switch between spinning up simple sites and more complex apps. I noticed my agents keep writing specs and testing in the browser, which is unnecessary. So I’m adopting this.

You don’t need to mention skills you’ve installed, as their ‘frontmatter’ (the skills’ name and description) is also pre-loaded alongside your AGENTS.md.

If you’re using ChatGPT/Claude Desktop apps these instructions still work. Paste them in ‘personalisation’ / ‘preferences’ / or ‘instructions’ - in your settings.

AGENTS.md also get dynamically loaded as an agent navigates through folders.

my-project/
├── AGENTS.md              ← root instructions (always loaded)
└── docs/
    └── AGENTS.md          ← loaded when agent works in /docs

If it was helpful, reply and let me know 😊

What am I building this week?

  • I’m currently in Cardiff doing talks with 16-18 yr olds on starting companies and AI. Making me think I should do proper workshops for these kids.

  • I just killed my email triage bot. Going to try a different approach. Tried Replit Agent 4, but just did not work. I’ll record a video on this build + Replit review

  • I’ve been using keep.md to bring all my saved items into one feed (that gets filtered for this newsletter) - and I’m going to make the feed public

  • I need to cut my AGENTS.md down + add conditional blocks (as above)

  • I may do a custom ‘claude for chrome’ extension…


Ben’s Bites is brought to you by Viktor

Your next hire isn’t on LinkedIn. Viktor252 is your last hire. He has his own computer, lives in your Slack, watches how your team works, and starts solving problems before anyone notices them. Writes code. Builds apps. Runs campaigns. No triggers. No workflows. It just works. Try Viktor free252, you have $100 up front.*


Headlines

  • Google launched the new Stitch456. They’re calling it your ‘vibe design partner’. Gemini is great at generating UI, unlike other models (ahem, openai), and now paired with their own harness with a bunch of features; AI-native canvas, design agent, voice, instant prototypes, design systems and DESIGN.md (handily links with our intro! Take a look at it257 😊 ). Here’s301 a good overview video of the tool.

  • Claude launched ‘Dispatch’ for Cowork213. Once you connect, you can send messages from your mobile to the Claude Desktop app (i.e. work with files on your desktop). It can also launch Claude Code sessions86. I got this working, but I had to approve some permissions on my Mac first (“git” & “claude” would like to access your files). They also published a report: What 81k people want from AI189

  • GPT-5.4 mini75 is available in ChatGPT, Codex, and the API75 (5.4 nano available in the API only). Roughly Sonnet 4.6 intelligence, 3x faster and ~70% cheaper. OpenAI also launched a challenge62 to build the most efficient pretrained model under extreme constraints.

  • Midjourney’s v8 model99 is in early testing. It’s 5x faster and better at rendering text, but nowhere near nano-banana. It wins in the aesthetics as always, though.

  • Runway’s new real-time video generation model120 created instant HD videos on Nvidia’s Vera Rubin hardware with time-to-first-frame under 100ms.


My feed


Afters

We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.

@brian_scanlan169

This was asked for for YEARS and I could never find time to build it myself 🗺️ Hoodmaps for 🏡 Airbnb Hoodmaps is my app that lets you find out where to stay in a city, it classifies neighborhoods by: 🟥 Tourists 🟨 Cool 🟩 Rich 🟦 Suits ⬜️ Normies I asked Claude Code to

@levelsio215

Today, we’re introducing Forge, a system for enterprises to build frontier-grade AI models grounded in their proprietary knowledge. 🌎 Forge bridges the gap between generic AI and enterprise-specific needs. Instead of relying on broad, public data, organizations can train models

@MistralAI127

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents

@MiniMax_AI83

Introducing MCP for arXiv Let your research agents stand on the shoulders of giants Fast multi-turn retrieval, keyword search, and embedding search tools across millions of arXiv papers 🚀

@askalphaxiv105

Introducing community plugins on Cursor Directory. Add directly via a GitHub link, we auto-detect rules, MCP servers, agents, skills, and more. Or create one manually. Following the Open Plugins standard.

@pontusab109


* sponsors who make this newsletter possible :)
Wanna partner with us for the next quarter?
Email us at shanice@bensbites.com28 or k@bensbites.com24