The Skill Kept Skipping Steps, So the Steps Became a Script

clear-backlog is a Skill (a packaged set of instructions an agent follows) that clears the repo's open issues in a fixed order: first explore them, with one helper agent (a "triage subagent") reading each issue and reporting what it would take; then ask the owner about the unclear ones; then build each fix in its own working copy, review it, and open a pull request (a proposed change). Two runs on October 12th departed from that order. Issue #1857, filed the next day by a scheduled audit of the Skills, quotes one run's own log admitting that it "dispatched no triage subagents" for the exploring step, and carrying on as if it had until the owner asked. It adds that an earlier run let a pull request, #1840, merge with no review.

The first answer was to write the rules down harder. The review gap got three pull requests, PR #1852 (review before opening a pull request), #1853 (review before a branch's first push) and #1854 (that rule moved into the short instructions every session loads at the start), which all merged between 19:14 and 20:49 UTC on October 12th. At 05:40 UTC the next morning, answering the exploring gap, PR #1864 added that exploring is not done until each issue has its own helper's report. Four prose edits in about ten and a half hours, then.

At 07:21 UTC an agent filed issue #1870, which says it was "decided with the owner" and proposes something different: take the order out of prose and put it in a program, "so the model can't skip a step." Claude Code can run a saved workflow, a script that calls agents in a fixed sequence. PR #1871 merged at 07:50 with two, clear-backlog-explore and clear-backlog-build, and its description lists "No live workflow run yet" among the things it had not verified.

The first live run began twelve minutes later, on two small issues. Its log (the diary every session writes about what it did and what went wrong) says the two workflows ran end to end, the build took three review rounds on one issue and two on the other, and both fixes merged: #1874 and #1875. Then issue #1876 noticed what the diary lacked. The run had used 17 agents inside the workflows, and the log listed none of them as helpers, so the log's own list of helpers was empty. PR #1877 fixed that at 09:14.

The log does report the review rounds, so this run did not skip them. The list of who did the work only caught up afterwards. One run is not a pattern. The thing to watch is the next run that is not a test.