An agent skill -- adaptogen -- that improves any skill
(itself included) by running it, adversarially critiquing the result, and
landing every evidence-backed fix immediately as a real edit, confirmed by
re-running and committed to git. No side files, no state tracking: git is the
entire durability layer.
Copy all of skills/adaptogen*/ (the adaptogen router plus its six
sibling state-skills: adaptogen-orient, adaptogen-run,
adaptogen-critique, adaptogen-land, adaptogen-commit-confirm,
adaptogen-build-new) into your project's own skills/ (or your
agent's skills directory, e.g. .claude/skills/) alongside whatever
skills you already have. The router dispatches the siblings by name, so
all seven must be installed together -- copying only skills/adaptogen/
leaves every transition dead-ending on a missing skill.
Point it at any skills/<name>/SKILL.md, including its own. The
adaptogen router walks a DAG of sibling state-skills: adaptogen-orient
classifies the target (existing vs new, git co-location), adaptogen-run
dispatches an isolated subagent to actually run it (or adaptogen-build-new
drafts one first), adaptogen-critique sends a second isolated subagent to
hunt for fault, adaptogen-land edits the file each confirmed fault is in,
and adaptogen-commit-confirm commits and re-runs to make sure the fix held
(reverting on regression). See skills/adaptogen/SKILL.md for the state map
and the exact procedure.
Checkpointing, rollback, and a record of what changed and why are all things a
git repo already does for any tracked file -- a commit is the checkpoint,
git revert is the rollback, the commit message is the record. Building a
second, parallel state-tracking layer on top of that would just be a worse git.