Give It Hands

Session 5 · Tools + Choosing the Approach · TCE · 50 min · pairs · Same rhythm.

Runs entirely in your browser (Google Colab). Any laptop or lab PC works — nothing to install, no GPU, no payment. See ZERO-SETUP.

Part A — Calculator tool (12 min)

Define calculator as a plain Python function with a real docstring. Hand it to the model. Ask the GST question — exact answer. Then ask without the tool and watch it guess. That contrast is the lesson: the model delegates instead of predicting digits.

✓ Checkpoint 1: with-tool exact, no-tool wobbles.

Part B — Two tools, chained (12 min)

Add days_between. Ask something needing both (the internship question). The agent loop chains them — nobody scripted the order.

✓ Checkpoint 2: both tools called in one answer.

Part C — See the machinery (8 min)

Disable automatic calling; print the raw function_call the model emits. This is the request; your code decides whether to run it. That's why the model can't actually do anything dangerous by itself.

Part D — Scenario cards (10 min, paper + pen)

Four scenarios → pick Prompt / RAG / Tools / Fine-tune, justify in 2 sentences each. A mini design doc — and capstone rehearsal.

✓ Checkpoint 3: defend one choice to me out loud.

Stretch

Plug your Session 4 search_notes in as a tool — your capstone becomes a real assistant (knowledge + hands) · manual agent loop · break-it-then-fix-the-docstring (proves docstrings are prompts).

Capstone

You now hold all three pieces: RAG (knowledge) + tools (hands) + evals (judge). Save the notebook. The finale attacks it, you harden it, then demo. Bring it charged.