Treat the 200-line plan as source of truth and review research before code — errors cascade from the top

HumanLayer Blog·medium signal

ACE-FCA's Research→Plan→Implement loop puts the highest-leverage human review on the research output, not the diff: 'a bad line of research lands you thousands of bad lines of code,' while a bad code line stays isolated. Reviewers read a ~200-line plan instead of 2,000-line PRs to keep mental alignment. One practitioner shipped 35k LOC into a 300k-LOC Rust codebase in 7 hours using this exact checkpoint discipline.