The system prompts that coding agent harnesses pass to language models are massive. They describe every available tool in detail — even the ones you never use.

So I wondered, what if I built something more minimal?

I tend to work on small projects where all the code fits in the context window. So I built a harness with a single tool: the ability for an LLM to edit my files.

The core algorithm:

1. Read every file in the project directory

2. Concatenate them with the user's prompt into a single message

3. Send it to any language model, instructed to respond with a message and a list of edits — each edit being a search and replace on a specific file

4. Apply the edits

I tried getting the LLM to generate unified diffs, but it kept producing malformed ones. Search and replace turned out to be far more reliable — the model copies text verbatim from the files already in its context.

This can be extended with a few more features:

1. A CLI that lets you continue the conversation or reset it. Dumber LLMs need this so it can see the mistake they made.

2. Automatic retries — if an edit fails, re-send the current files and have the LLM fix its own search blocks.

3. File deletion. I made it so that an empty search-and-replace on an empty file deletes it.

The LLMs I tested were local ones running through the LM Studio API. Gemma 4 was pretty bad. Qwen 3.5 was significantly better.

Where is the source code? You should be able to get any coding agent to generate you one using this post.