Kushal Chakrabarti studies reliable AI and advises builders at South Park Commons. A former chief research officer at Opendoor, he writes Obviously Wrong and researches how coding-agent instructions accumulate and can be maintained.[1][2][3]
Start with Why Does CLAUDE.md Keep Growing? →
X Career
South Park Commons describes Chakrabarti as an advisor and former chief research officer at Opendoor. It also credits him with founding two AI and fintech startups and contributing to Amazon’s early personalization engines. His February 17, 2025 departure essay describes five years at Opendoor, beginning as an advisor before joining full time. It recounts adapting the research organization through the company’s expansion and the subsequent change in housing-market conditions.[1][4]
Contributions
His paper on catastrophic remembering studies coding-agent instruction files across 1,867 repositories. Appending a rule is easy; safely deleting it becomes difficult when the reason for adding it has disappeared. He reports that prompts more than tripled over their observed lifetimes. The proposed intervention is to keep a rule’s rationale alongside it, so later changes can distinguish a necessary constraint from an obsolete workaround.[3]
Ideas
Measure maintenance, not just prompt length
The paper tests rationale comments in constructed environments with known optimal prompts and on instruction-following tasks. Those experiments support preserving reasons in the tested tasks. They do not establish that every repository can achieve the same reduction. Record which failure each rule addresses.[3]
Writing about reliable AI
Obviously Wrong is Chakrabarti’s publication on AI, leadership and technology. He describes his recent focus as efficient, reliable AI and draws on experience in personalization, marketplaces and applied modeling. The writing connects technical choices with the organizational work of deploying models that influence consequential business decisions.[2]
Tools, writing & talks
Writing, research & talks
Start here
Why Does CLAUDE.md Keep Growing?arxiv.orgA study of growing agent instructions and experiments that preserve the reasons behind rules.
Read the abstract and experiment setup to decide whether rationale comments could help maintain your own agent instructions. Familiarity with coding-agent rule files is useful; distinguish the repository observations from the controlled tests.
Sole listed author[3]A study of 1,867 repositories finds agent instruction files like CLAUDE.md more than triple in size because rule rationale is lost. Commenting each rule with its failure and evidence kept prompts near ideal length with equal constraint adherence.
Catalog workOriginal source- Working Backwards & Forwardsobviouslywrong.org
A first-person account of research leadership through five years at Opendoor.
Author[4]