The New Foundation for Researchers
Pick tools from first principles instead of habit, and know which tasks AI should not touch.
Before this:00 · Orientation: What You Have
The tool list in this track will be stale in eighteen months; the judgment in this unit will not. You already know how to evaluate methods and distrust convenient results, and that is the old foundation: AI-assisted work spends that muscle faster, not less. So before any unit about tools, this one, where you decide what the machines get to touch.
Learning outcomes
- Define vibe coding on this curriculum's terms, and say what it is not.
- Distinguish a workflow from an agent, and explain why the distinction changes how you debug.
- Select a tool for a task from the task's shape rather than from habit or novelty.
- Identify at least one task in your current coursework where AI is the wrong instrument, and defend the call.
What you already have
Five surfaces, each good at something different. The guide's own summary:
| Task | Tool |
|---|---|
| Chat-based research, writing, summarising documents | Gemini Advanced (browser) |
| Searching your own Drive files and email | Gemini + @Drive / @Gmail |
| File operations, shell scripts, code generation in a terminal | gemini-cli |
| Interactive data analysis, model training, Python notebooks | Google Colab AI |
| Multi-agent coding, full codebase navigation | Google Antigravity (preview) |
The table tells you which tool. It does not tell you whether to reach for one at all.
Concept
Requirements, specs, structured prompts, agents: that is the stack, and notice that no product name appears anywhere in it. The tools underneath rotate every quarter, which is exactly why the stack is the thing worth learning.
Part 1 is four short sections and worth reading in order: §1.1 (Why the old foundation matters MORE now), §1.2 (What "vibe coding" actually means), §1.3 (The new skills stack), and §1.4 (When NOT to use AI to write your code).
§1.4 is the one to read twice. It carries the AFD test: the category of code where being wrong hurts someone who is not you. Graduate research work crosses into it more often than most students expect.
Paired instantiation
| Portable idea | Google (what you have) | Internet Menace (the method) |
|---|---|---|
| Match the tool to the task shape | The task→tool table | §1.3: the skills stack underneath it |
| Know what the tool cannot do | Preview caveats, rate limits | §1.4: when not to use AI at all |
| Name the thing you are doing | "Chat", "agent", "notebook" | §1.2: vibe coding, defined with teeth |
Lab
- List three tasks you are actually doing this term. Be specific: "clean the 40k-row survey export", not "data work".
- Route each through the task→tool table above and write down the tool it points at.
- Now read §1.4 and apply the AFD test to each. Does being wrong cost you an hour, or does it cost somebody else something they cannot get back?
- For any task that fails the AFD test, write down what you would hand to a person with credentials and liability, and at what point.
- For the remaining tasks, write one sentence on why the tool the table picked is right, or why it is wrong and what you would use instead.
Deliverable
A one-page tool-selection rationale covering three real tasks from your current coursework: the tool chosen, the reason, and the AFD verdict for each.
Self-check
- What makes a program an agent rather than a workflow, and why does that change how you debug it?
- You have a task where the output is non-deterministic and the grader expects one answer. What does that tell you?
- Which of your three tasks would you refuse to hand to an agent even if it produced a plausible result, and why?
- The task→tool table has five rows. What is it missing?