AI-Assisted Workflows · Research

AI-Assisted Development Exercise


As of this writing (2026). This tool landscape changes fast — treat the table below as a snapshot, not a current recommendation, and revisit it periodically.

Definition

This is the applied, current-tools version of Benchmarking a New Tool: pick an AI coding assistant, use it to build a solitaire simulator, and produce an evidence-based reflection on the experience — before you have to rely on an AI coding tool for something that actually matters, like a real research demo.

Learning Outcome

After completing this exercise, you should have hands-on experience with at least one current AI coding assistant on a low-stakes, checkable task, and a documented, specific sense of what it’s good and bad at — rather than a vague general impression of “AI coding tools.”

Core Structure

The benchmark tool landscape, as of 2026 (each entry: what it’s good for in this context):

Tool IDE Compatibility Why It’s Good Here
GitHub Copilot VS Code, JetBrains, Neovim, Visual Studio Industry-standard; strong completions, refactoring, and tests while keeping the student responsible for correctness
Claude Code VS Code (via extension), web Strong at reasoning about game logic and state transitions; explanatory, conversational, promotes actual learning
Replit Browser-based IDE Very low setup friction; good for rapid prototyping and sharing a playable version
Amazon Q Developer VS Code, JetBrains, AWS Cloud9 Exposure to enterprise-style AI assistance and secure-coding hints
Zed Zed Editor (macOS, Linux) Good for tooling- and performance-focused students; modern editor design with emerging AI features
Tabnine VS Code, JetBrains, Eclipse, others Solid autocomplete with a privacy focus; less conversational
AntiGravity Experimental standalone IDE Useful for discussing intent-driven programming paradigms; not mature enough for full projects yet
OpenAI Codex API-based, no IDE Less suited as a primary learning tool here; obscures the human-in-the-loop workflow this exercise is meant to build

The exercise itself:

  1. Build a solitaire simulator using one or more of the tools above.
  2. Play at least 1000 simulated games and compute the win rate (wins ÷ games played).
  3. Write a short report (roughly two pages) covering: which tool(s) you used, your experience with them (both positive and negative, specifically), and your winning percentage with the details of how you computed it.

Working alone or in a group is fine, but in a group, every person needs their own code, results, and report — the point is individual hands-on experience with the tool, not a shared output.

Worked Example

A student picks Claude Code for its conversational, explanatory style. Early on, the tool correctly implements the core deal-and-draw logic but needs a follow-up question to get the three-pass-through-the-deck rule right — a concrete, specific data point for the report, not just “it mostly worked.” After 1000 simulated games, the win rate comes out in a plausible range, which is itself a useful sanity check: an implausible result (near 0% or near 100%) would be a signal that either the game logic or the tool’s implementation of it has a bug, caught cheaply because the benchmark’s checkable.

Common Pitfalls

  • Running too few trials to get a stable estimate — a win rate from 20 games is much noisier than one from 1000.
  • Reporting only whether the tool “worked,” without the specific detail (what needed a follow-up, what it got right immediately) that actually builds calibrated judgment.
  • Treating an implausible result as acceptable rather than as a signal to debug the implementation.
  • Skipping ahead to using an AI tool on the real project without ever having exercised it on something checkable first.

Checklist

  • At least one current AI coding tool used hands-on
  • 1000+ simulated games run to produce a stable win-rate estimate
  • Report documents specific positive and negative experiences, not just a verdict
  • Result checked for plausibility, not just reported
  • If working in a group, each person has their own code, results, and report