ADD · AI-Driven Development
Measured on Claude Sonnet 5.5 · October 2026
Same AI. Same 300 real bugs.
One difference: ADD.
Claude Code fixed 300 real open-source bugs twice: on its own, and with ADD.
Claude Code on its own
Claude Code + ADD
THE IDEA
Why write the plan and the tests first?
Straight to code
Request→Code→“Done”
With ADD: every step leaves something you can check
Request→Plan→Failing test→Code→Passing test→Evidence
A silent guess gets written down
A test written first proves the fix
A saved check can't be quietly weakened
The next session can read it
1 · BUGS FIXED
Likely real, not proven.
2 · CHECKED BEFORE “DONE”
ADD ran the tests and saw them pass on almost every fix.
One dot per bug. Filled: a passing test run was seen before the fix was handed back.
3 · SHIPPED WITH A TEST
Almost every ADD fix carries its own test.
4 · WHEN THE REQUEST HAS A GAP
Nobody said who may cancel a booking. ADD chose owner-only and wrote it down.
Runs that kept cancelling to the owner, out of 3.
Planted bugs caught by the AI's own tests.
5 · THE PRICE
ADD costs more:
6 · ONE SETTING MATTERS
Run ADD at medium effort. At low it skipped its own instructions half the time.
claude --effort medium
7 · ONE REVIEWER, 6 BUGS, BLIND
The reviewer preferred the shorter output every time. Approving it was less safe.
Which output the reviewer preferred.
Reviewer's verdict agreed with the benchmark's tests. One reviewer, not timed: a hint, not a measurement.
NOT MEASURED YET
Timed review · brand-new bugs · repeat runs
Review time is not measured. These 300 bugs are public and older than the AI. Each side fixed each bug once.