docs: improve readme narrative and setup

Author: lvillani

README.md

@@ -18,18 +18,45 @@
 
 ## 🧭 Why this exists
 
-`thing` began as a question: how does an agent harness actually work? Rather than trust
-a framework's magic, I hand-wrote just enough code to make a single API request, then
-let the harness pull itself up by its own bootstraps. The working prototype gained the
-`bash` tool by having its human-in-the-loop upload its own source code, fed back in
-base64-encoded chunks. Once it could see and edit its own code, I started adding the
-features I enjoy from other harnesses, borrowing most heavily from [Pi](https://pi.dev),
-which I use every day.
+`thing` started with a simple question: how does an agent harness actually work? I
+hand-wrote the first loop, then uploaded its source code to the running agent in
+base64-encoded chunks. I asked it to implement the `bash` tool and copied its answer
+back into the source. That was the moment the harness began to evolve: it could now
+inspect and edit its own code.
 
-No vibe-coding. The whole point is to understand how agent loops work. A model may help
-guide its own evolution, but nothing ever lands blindly.
+From there, I began adding the features I enjoy in other harnesses, borrowing most
+heavily from [Pi](https://pi.dev), which I use every day.
+
+No vibe-coding. The goal is to understand how agent loops work. A model can help guide
+its own evolution, but nothing lands blindly.
+
+## 🧰 Getting started
+
+`thing` connects to an OpenAI-compatible Chat Completions endpoint. Create
+`~/.config/thing/config.toml` with your model and endpoint:
+
+```toml
+model = "your-model"
+endpoint = "https://your-provider.example/v1/chat/completions"
+reasoning_effort = "medium" # optional
+```
+
+Then run:
+
+```sh
+./script/run
+```
+
+On first run, `thing` asks for an API token and stores it in the system keychain.
+
+## ⚠️ Current trust model
+
+Tool calls are not confirmed one by one, and sandboxing is not implemented yet. The
+`bash` tool runs with the permissions of the process, so use `thing` only in a project
+environment where you trust the model and can tolerate its changes.
 
 ## 📚 The paper trail
 
-The `docs/adr/` directory records *why* decisions were made. If you want to understand
-the trade-offs, that's where to look.
+The `docs/adr/` directory captures the reasoning behind the project. It records the
+important decisions, the alternatives considered, and the trade-offs that shaped the
+harness.