A developer shares practical lessons from a month of working with GitHub Copilot on a personal project. Key findings: Copilot works best with small, well-scoped tasks rather than large end-to-end requests; context window limitations cause it to forget earlier instructions or fixate on unwanted patterns; terminal access and refactoring are weak spots compared to traditional IDEs; model differences are less impactful than task size; unit tests passing doesn't guarantee correct behavior; and spec documents need to be scoped per-task and discarded after use to avoid polluting context. The author recommends an incremental 'plan, review, implement' approach and verifying an MVP early.
Table of contents
Copilot is very eagerContext mattersAccessing the terminal is an unnecessary painModels are sodaAI is like an unreliable teleportation deviceUnit tests come with limited warrantyDesign documents are yet another thing that require maintenanceConclusionSort: