💬 The Gemma 4 Playbook

The two constraints, decoded

• 2000 output tokens ≈ ~120 lines of code or ~1200 words. Long answers get cut off mid-line. So never ask for "the whole solution" — ask for pieces.
• Unlimited new chats = unlimited retries with a clean slate. Fresh chats are free. Use them aggressively; never argue with a confused model in a long thread.

Core rules

1. Chunk everything. Ask for ONLY the Dataset class → ONLY the model → ONLY the training loop → ONLY the submission writer. Magic phrases: "Code only, no explanation." · "Max 30 lines." · "Continue exactly from this line: …" (when cut off).
2. One chat = one job. Explainer chat, bug-fixer chat, strategy chat, API-lookup chat. Contexts stay clean, answers stay sharp.
3. Paste context, always. Gemma can't see your screen. Every prompt: the task in 1–2 sentences, the smallest relevant code, exact error/traceback, tensor shapes, one specific question.
4. Verify everything. Read the code before running (does it use variables that exist?) → run on a tiny slice → print(x.shape) liberally → if a fix fails twice, fresh chat with the new traceback.
5. You are the manager. You choose the plan (the 8-step flow); Gemma writes boilerplate, fixes bugs, explains baselines, and recalls API signatures.

Contest-day prompt bank

“Summarize this task statement in 5 bullets: input, output, metric, submission format: <paste>”
“Baseline does X and scores Y. List 5 improvements ordered by (impact ÷ time) for a 6-hour contest.”
“Explain what this code does, line by line, for a beginner: <paste baseline chunk>”
“This code raises this error. Code + full traceback below. Give the minimal fix only.”
“Rewrite this training loop to add early stopping on validation loss. Change nothing else.”
“My model overfits (train 0.99 / val 0.70). The 3 fastest fixes for a CNN in torch?”
“Write code to save predictions as CSV, one 0/1 per line, no header, from tensor preds.”
“How do I load a pre-cached Hugging Face model from a local path, no internet?”

Time economics (6 hours, 3 tasks)

• First 5 min: read all 3 statements; rank by expected points-per-hour.
• First 30 min per task: baseline running and submitted.
• While models train: a Gemma chat plans the next improvement in parallel.
• Last 45 min: freeze experiments; verify formats; confirm your best submissions are the selected ones.