Cap generated keywords and simplify LLM correction retries.

AGENTS.md

@@ -1,27 +1,42 @@
 # Agent Instructions: PhotoTools3
 
 ## Environment Setup
+
 Requires a `.env` file with the following variables:
+
 - `llm_api_key`: API key for the LLM provider.
 - `llm_api_base_url`: Base URL for the LLM API.
 - `llm_default_temperature`: Float value for temperature.
 
 ## System Dependencies
+
 - `exiftool` binary: required by `pyexiftool` at runtime, not installable via pip/uv.
   - Arch: `perl-image-exiftool`
   - Debian/Ubuntu: `libimage-exiftool-perl`
 
 ## Development Toolchain
+
 - **Dependency Management**: Use `uv`.
 - **Python Version**: Requires Python `>=3.14`.
 - **Testing**: Run `pytest` to execute the test suite in `test/`.
 - **Linting & Formatting**: Use `ruff` and `black`.
 
 ## CLI Usage
+
 The primary entry point is the `photools3` command:
+
 - `photools3 keywords <image_path>`: Extracts keywords and captions from a single image.
 
 ## Architecture & Logic
+
 - **`src/photools3/photo.py`**: Core image processing logic, including resizing, RGB conversion, and generating base64 JPEG encodings specifically for LLM vision compatibility.
 - **`src/photools3/llm_session.py`**: Manages the workflow of sending processed images to the LLM and parsing results.
 - **Configuration**: The list of keywords to extract is maintained in `config/keywords.txt`.
+
+## Code style
+
+- Avoid overly verbose docstrings or comments. Be precise!
+
+## Comunication
+
+- When done with code modifications, give a one-liner for the currently unstaged changes that can be used as a commit message.