I Tried to Recreate a 2008 .NET Developer With AI. I Broke My Own Experiment 4 Times.

Description and use

The experiment demonstrated that AI can produce modern code but struggles to replicate 2008 practices. Separating legacy generation from modernization yields a clearer migration workflow, mirroring real-world scenarios.

Problem

Determine if AI can generate code that reflects the practices, constraints, and style of a 2008 .NET developer in a real legacy environment.

Outcome

The experiment showed that AI can generate modern code but struggles to replicate 2008‑style practices without explicit constraints. Separating legacy generation from modernization yields a clearer migration workflow, mirroring real-world scenarios where the original developer and the modernizer are different people.

Workflow

• 1. Define Yamada’s background, company context, and 2008 development environment.
• 2. Prompt Claude Code to generate a 2006 equipment tracker reflecting Yamada’s style.
• 3. Analyze the generated code for 2008 characteristics (Option Strict Off, Windows Forms, etc.).
• 4. Use the legacy code as a baseline for a new sales system in Phase 01.
• 5. Attempt to generate 2008‑style code with Codex; iterate prompts and instrument setups four times.
• 6. Observe Codex consistently produces 2026‑style code, indicating difficulty in reproducing 2008 style.
• 7. Decide to separate legacy generation (Claude) from modernization (Codex) to isolate variables.
• 8. Modernize the legacy code with Codex, demonstrating a realistic migration workflow.
• 9. Document lessons learned about instrument consistency and the separation of legacy and modernization tasks.

Evidence

"Write some old-looking VB.NET from around 2008"

Source

"Codex can't write 2008."

Source

"Drop the comparison."

Source