About this video
Your AI agent is failing because you are feeding it too much garbage. In this video, we explore why bloated Claude MD files are the enemy of productivity and how to reclaim control. We dive into Anthropic's new interactive initialisation process and show you a secret method to override the default settings with your own custom skills. Key takeaways from the video: - Why minimal files prevent agent confusion and performance bloat. - How to enable the hidden interactive initialisation setting in your json config. - The difference between Project and Local permissions for shared workflows. - Creating a custom init skill to automate project setup. - Using the agents.md standard to ensure your configuration works across Claude, Zed, and Kilo. Sources: https://arxiv.org/pdf/2602.11988 https://arxiv.org/pdf/2602.12670 https://x.com/trq212/status/2035799806640115806
Mastering Claude Initialisation: Stop the Bloat and Start Coding
Most developers are sabotaging their AI agents without even realising it. By stuffing Claude MD files with every minor instruction and preference, we create a landscape of confusion and unnecessary bloat that leads to the agent ignoring critical commands. To get the best out of Claude, you need to stop micromanaging and start orchestrating.
The Problem with Bloated Instructions
There is a growing controversy regarding how MD files are impacting agent performance. When a file becomes too large, the agent prioritises the wrong information or simply gives up on following complex instructions. The solution is simple: keep your files minimal. You should only nudge the agent in specific directions, rather than trying to map out every possible behaviour.
Enabling the New Interactive Initialisation
Anthropic has introduced an experimental interactive process for setting up projects. To access this, you must modify your settings.json file. By setting claudecodeinit to true, you unlock an interactive prompt that allows you to define permissions and skills at the start of every project. This ensures that permissions are handled correctly, whether you are working on a local-only basis or sharing configurations across a team project.
Taking Control with Custom Init Skills
While the default initialisation is a great step forward, the real power lies in customisation. You can actually override the default /init skill by creating your own skill markdown file. This allows you to define a bespoke setup process that fits your specific workflow, such as using Bun instead of NPM or automatically setting up complex folder structures.
The Agents MD Standard
One of the most effective ways to manage your AI environment is to adopt the agents.md standard. This is a format understood by other major agents like Zed and Kilo. By symlinking your claude.md to an agents.md file, you create a universal source of truth for your project rules. This not only keeps your project clean but ensures that no matter which AI tool you use, the rules remain consistent and efficient. Stop feeding the bloat and start building with precision.