Superpowers 7.0.1 keeps its five-round fix limit after Claude Code compacts
Superpowers 7.0.1, released October 21, shrinks its subagent-driven-development skill under twenty thousand characters, so Claude Code keeps the skill's five-round fix limit after a compaction. In a long session, that limit is what stops one task from looping on fixes.
After a compaction, Claude Code re-attaches each skill cut at 20,000 characters
After a compaction, Claude Code re-attaches each skill you invoked, cut off at twenty thousand characters. subagent-driven-development ran past thirty-two thousand. So the fix loop, its five-round cap and the final review were gone for the rest of the session.
One task ran 18 fix rounds before its human stopped it
One reported session ran about thirty-three hours, through thirty-six compactions. The skill allows five fix rounds per task. One task ran eighteen, and had started a nineteenth, before the human stopped the loop.
Three skills now fit under the cut
The fix moves the fix loop, the final review and five other steps into reference files, which the skill reads from disk at the step that needs them. writing-skills splits the same way, and executing-plans drops a repeated example.
A new test fails when any SKILL.md reaches 20,000 bytes
A new test fails whenever any skill file reaches twenty thousand bytes, so the next addition can't push one over. It can't help when Claude Code drops a skill from the re-attached list altogether, or when rules break before any compaction at all.
Brainstorming writes a design doc before it builds, even for a small change
Brainstorming changes too. It now lays out two or three ways to build the thing, recommends one, and asks you to pick. Then it writes a design doc, a short one for a small change, and builds nothing until you approve it.
Skipping the design doc without asking fell from 24 of 24 runs to 3
Asked for a small change to a photo gallery, the old skill skipped the design doc without asking in every run. Now it skips in about one run of eight.
Brainstorming still waits to be asked for approaches when you keep asking for a recommendation
Still weak: keep asking “what would you recommend?” and approaches often wait until asked
Spikes and an explicit “just build it” skip the steps. The compaction fix targets Claude Code.
One weak spot remains. If you keep asking what it would recommend, brainstorming often lays out approaches only when you ask for them. Throwaway spikes, and saying just build it, still skip the steps. How many compactions does your longest Claude Code session run through?
