Shreya Shankar routes coding tasks by size and separates bug review
Shreya Shankar describes a coding workflow built around human-reviewed RFCs, implementation routing by task size and separate bug review. Follow-ups explain section-by-section specs and restarting failed approaches.
TL;DR
- Implementation routes by expected size: Shreya Shankar, a computer science researcher, uses Fable above 4,000 lines and Opus 5.5 otherwise in her workflow.
- Bug review and simplification have separate owners: Astra finds bugs; humans propose code reductions that Astra executes in the same pipeline.
- RFCs require a complete human read: Shankar commits to understanding every word, she explains in a follow-up.
She builds RFCs outline first, then section by section. Her best detail is the restart rule: a bad approach sends both the RFC and its implementation to the bin.
The 4,000-line cutoff
Shankar describes six stages, with the implementation model selected by how much code she expects the RFC to require:
- Issue: A human creates the issue or feature idea.
- RFC: A human iterates with Opus 5.5 on the specification.
- Implementation: Fable handles work expected to require more than 4,000 lines of code; Opus 5.5 handles the rest.
- Bug review: Astra reviews PRs for bugs and re-invokes the implementation agent.
- Simplification: A human reviews the PR code and proposes simplifications; Astra carries them out.
- Documentation: A human works with Sol.
RFCs, outline first
Prompt skills and keywords help keep the RFCs concise but comprehensive. Shankar ties that constraint to her own reading burden: “If I have to read everything I know I’ll make sure these RFCs are efficient.”
Bug review and code deletion
The human simplification pass has an explicit objective: minimize lines added and maximize lines deleted. Astra executes those changes after the human proposes them.
Asked whether the review catches “drift,” Shankar keeps the target at bugs. She calls drift ill-defined and treats bugs as one form of it.
Discarded RFCs and PRs
Once implementation exposes a bad approach, Shankar abandons both artifacts and starts a new RFC describing a different approach, rather than patching the existing specification.
Manual branch handoffs
Shankar calls herself the “meat proxy” between agents. She starts new sessions and asks the AIs to inspect a particular branch and perform a specified task.