Build neural network from scratch, one small file at a time

Features

• Filesread_file, write_file, edit_file, move_file, copy_file, list_dir, grep and glob work in the chat's working folder; a counter shows files created, edited and read.
• Shellruns terminal commands such as your training command and tests, in Windows, WSL or over SSH, and reports the real exit code. Dangerous commands are blocked.
• Webweb_search searches the web and http_get reads one documentation page, so the agent can look up the meaning of an error message.
• Screenshots and imagesgen_image saves real image files, and ocr reads text from a saved screenshot, for example the numbers on a loss chart you captured.
• Orchestrationtask runs parallel sub-agents, run_on_model sends a hard step to a bigger model, and skill stores a reusable playbook you can run again.
• Approval modesAuto-run acts at once, Ask first pauses before actions that change something and shows the diff to accept, and Plan only writes a plan and runs nothing.
• Cronruns a prompt on a schedule, for example to rerun a small experiment each night and write its loss to a CSV file you read the next day.
• Modelsconnect a cloud model through a standard API endpoint or a model on your own PC, or use the built-in model, which is ready without your own endpoint.

Walkthroughs

Write a perceptron in plain Python

Understanding the perceptron is the first step when you want to build neural networks from scratch in Python. You ask for a small file, and write_file creates a plain Python file with a tiny list of examples and one neuron that learns the AND gate with a fixed learning rate in a short loop of weight updates. Shell runs it on your PC and prints the weights in the chat card, together with the real exit code. If you need help with your first neural network, start with one neuron. No framework is needed, and the file is short enough to read in one go. In Ask first mode you read the new file in the card and accept it before it is saved, then you check the printed weights yourself.

Find and reduce overfitting

Overfitting shows up when training loss keeps falling while loss on separate validation data stops improving or rises. grep searches your log for both loss columns, and read_file opens the training script so the agent can see the model size and any regularisation settings. Common remedies are more data, a smaller model, regularisation or early stopping. Keep each experiment small and repeatable, so a change in the numbers has one clear cause. With edit_file you get a diff to review, and in Ask first mode nothing is saved before you accept it. Shell then reruns your test command and returns the exit code, and you compare validation loss before and after the change on data the model has not seen, with the test set left untouched until the end.

Fix an input shape error

An input shape error means the data fed to the network does not match the first layer, so your script stops with an error message that names the 2 shapes it could not match. The input shape for neural networks has to match the data, and the output shape has to match the number of labels. The agent opens your neural network code in Python with read_file, uses grep to find the layer definitions and the place where batches are reshaped, and proposes a one-line change with edit_file. Plan only runs nothing, so use it first to read the explanation. Then switch to Ask first, accept the diff, let Shell rerun the script and read the printed shapes yourself instead of trusting the exit code alone.

Log training metrics to a CSV file

To compare experiments you need the loss of every epoch in a file, not only on screen. write_file creates a loss log, or append_file adds one row after each epoch, and read_file locates the place in your script where the loss is calculated. edit_file adds the logging lines and shows a diff. Shell runs your training command and returns the real output and exit code in the chat card. Ask first pauses before the script is changed, so you can read the new lines in the card first. Open the CSV file afterwards, put 2 runs side by side and note which settings you changed before you keep the change. The file is plain text in your working folder, and read_file can load it again in the next chat to continue where you stopped.