diff --git a/README.md b/README.md
index e69de29..6105608 100644
--- a/README.md
+++ b/README.md
@@ -0,0 +1,67 @@
+# Tic Tac Toe
+
+A Tic Tac Toe implementation where two reinforcement-learning agents
+(temporal-difference / value-based learning) train against each other, and a
+human can play against a trained agent from the command line.
+
+## How it works
+
+- `state.py` — board representation (3x3 numpy array of `'X'`, `'O'`, `'-'`)
+  and helpers to detect potential moves and winning lines.
+- `a_player.py` — `APlayer` base class shared by all player types (move,
+  win-check, game lifecycle hooks).
+- `machine_player.py` — `MachinePlayer`, a TD-learning agent. It keeps a
+  value table (state string -> expected reward) in a dict, picks the move
+  leading to the best-valued next state, occasionally explores a random move
+  (`p_explore`), and updates values towards the outcome of each game using a
+  learning rate `alpha`.
+- `human_player.py` — `HumanPlayer`, lets a human enter moves via the
+  console (German prompts).
+- `player_provider.py` — `PlayerProvider` pairs two players for a game and
+  randomly decides who starts.
+- `helper.py` — small numeric helpers (softmax/sampling), not currently used
+  by the main game loop.
+- `tic_tac_toe.py` — entry point: runs training and/or play loops.
+- `values_x.json` / `values_o.json` — persisted value tables for the X and O
+  machine players, loaded/saved across runs so training progress carries
+  over.
+
+## Requirements
+
+- Python 3.12+
+- numpy (see `.venv` for an existing virtual environment, or install with
+  `pip install numpy`)
+
+## Usage
+
+Run from the project root:
+
+```bash
+python tic_tac_toe.py [options]
+```
+
+Options:
+
+| Flag        | Default | Description                                                     |
+| ----------- | ------- | --------------------------------------------------------------- |
+| `--verbose` | `False` | Print detailed per-move debug output                            |
+| `--train`   | `False` | Train the machine players before playing                        |
+| `--num`     | `5`     | Number of training iterations / batches                         |
+| `--human`   | `False` | Let a human play as X against the machine player as O           |
+| `--load`    | `True`  | Load existing value tables from `values_x.json`/`values_o.json` |
+
+Examples:
+
+```bash
+# Train the two machine players against each other, saving progress
+python tic_tac_toe.py --train --num 50
+
+# Play as a human (X) against the trained machine player (O)
+python tic_tac_toe.py --human
+
+# Watch two machine players play a few verbose games
+python tic_tac_toe.py --verbose
+```
+
+During training, `values_x.json` and `values_o.json` are updated after each
+batch of games, so training can be interrupted and resumed.