What they did
Souleymane Sidibe, who mostly builds mobile apps, describes a year of working with Claude Code in stages. He added a CLAUDE.md per repo and custom commands, then chained planning skills (a questioning session, a spec, a split into issues, then implementation) so each step reads a file instead of a long chat. He wrote a stricter implement skill that demands real command output as evidence, tried a shell loop that fed the agent issue after issue, and then a manager skill that ran one background session per ticket in waves that avoid touching the same files, with locks on the shared emulators. That setup led him to build Loop, a Kotlin harness that pulls issues from a tracker, gives each one its own workspace and agent, and ends every run with a verdict from checks the agent did not write. Agent pull requests get a risk score from a cheaper model, and only low-risk ones merge without him.
How it went
He ran the implement skill 138 times before building the manager. Measured cost per issue had a median of 5 to 18 million tokens, one issue in ten went past 45 million and the worst passed 100 million, mostly runs nothing stopped. In one run every check was green but the agent had left uncommitted files, so Loop refuted it and sent the issue back. Since its first commit on November 5, Loop has 643 commits, 96 merged pull requests and about 2,270 tests, mostly written by Claude Code, and it now works on its own issues. Per-issue token budgets are not built yet.
Worth knowing
The self-run loop and the manager skill were fragile: the same agent wrote the code and the checks that called it done, and nothing capped spend. Weakened tests, red CI or diffs over 1,000 lines always go to a person. Loop's repo is private for now, so the numbers are the author's own.