We don't want to under-invoke working-in-isolation, since it protects against annoying bad behavior, but in some cases (particularly the plan mode workflow), the agent is basically already sure of the isolation state before it even invokes the skill. Claude's plans will sometimes contain lines like "The next required skill is working-in-isolation. We're already on feature-branch, so the skill will just show how to ensure isolation".
I think the skill winds up being completely useless overhead in these cases, since there's really nothing to teach. This leads to working-in-isolation eating up a lot of tokens for little value - it's loaded in full, then it's only instruction is "skip this entirely".
I'd like to find some kind of escape hatch, to avoid at least reading the entirety of the skill when it will just be skipped. This could be as simple as just splitting all the relevant instructions into separate files, each only accessed when the condition is met. hardening-plans could also explicitly call out an escape hatch, so it can jump to test-driven-development or no skill next if isolation is already guaranteed. I think something like this risks under-invocation though - were worktrees checked for? Is the current branch the default branch with a unique name? The skill is meant to guard against edge cases like this.
We don't want to under-invoke
working-in-isolation, since it protects against annoying bad behavior, but in some cases (particularly the plan mode workflow), the agent is basically already sure of the isolation state before it even invokes the skill. Claude's plans will sometimes contain lines like "The next required skill isworking-in-isolation. We're already onfeature-branch, so the skill will just show how to ensure isolation".I think the skill winds up being completely useless overhead in these cases, since there's really nothing to teach. This leads to
working-in-isolationeating up a lot of tokens for little value - it's loaded in full, then it's only instruction is "skip this entirely".I'd like to find some kind of escape hatch, to avoid at least reading the entirety of the skill when it will just be skipped. This could be as simple as just splitting all the relevant instructions into separate files, each only accessed when the condition is met.
hardening-planscould also explicitly call out an escape hatch, so it can jump totest-driven-developmentor no skill next if isolation is already guaranteed. I think something like this risks under-invocation though - were worktrees checked for? Is the current branch the default branch with a unique name? The skill is meant to guard against edge cases like this.