Review requirements, generate code and tests, then cut low-value tests
At a glance
A Lobsters developer describes an LLM workflow that ends with test consolidation and human review before sharing code. This workflow is listed under Coding Agents, Testing, Code Review.
- Agents / tools
- LLM coding agent, Superpowers
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Use a coding agent with access to your repository and its test runner. Start by discussing the feature, intended behavior, and corner cases; the contributor mentions Superpowers as a way to encourage requirements gathering and TDD. 2) Ask for a high-level design document and review it before letting the agent implement both the feature and its tests. 3) Once that work is ready, explicitly ask the model to trim and consolidate the generated tests, keeping cases that establish useful behavior rather than verbose or redundant checks. 4) Personally review the implementation and tests. Remove remaining low-value cases, by hand or with the model, before sending the change to teammates for code review.
How it went
Contributor koreth reports removing about 40% of the original test code on one change without losing much utility.
Worth knowing
This is one developer's reported review process; other participants in the thread disagree about having a model generate both code and tests.
