Give optional automation an exit criterion
Use when
A slash command or prompt wrapper is repeatedly bypassed, endlessly tuned, or adds no demonstrated value beyond a direct request. Exclude security controls, required verification, and rare but important recovery procedures; low usage alone is not a retirement case.
Action
In the existing project record, write the helper’s trigger, actual use, promised benefit, dependencies, and rollback. For an explicitly authorized, budgeted pilot, compare fresh sessions with and without that helper, keeping tasks, context, model, tools, and safety policy fixed. For a bug-fixing wrapper, the alternative is the same issue and relevant context in a direct request—not less information. Predeclare the benefit required to justify retention: better accepted outcomes or lower human effort without quality loss. Record repair work, elapsed time, and usage. Vary only the optional helper; do not disable enforcement to make the baseline faster.
Acceptance check
Both conditions must pass the same independent task checks and required negative safety tests. Inspect traces for accidental helper loading; retain failures, not just successful examples. Keep the helper if it earns its predeclared benefit. If it does not, and the alternative preserves outcomes, propose reversible deactivation with source history intact. After approval, verify ordinary tasks, dependent workflows, and rollback. Inconclusive results remain inconclusive. Never delete live dependencies or mandatory controls as “unused.”
Evidence
Ronacher names failed commands, explains why his bug-fixing wrapper added little, and publishes an acceptability-based repeat-run procedure. His later Pi account still discards unneeded skills while retaining useful review, todo, and interception extensions. The lesson is selective retention, not “automation never works.” Anthropic’s evaluation guide supports repeated trials and grading actual outcomes.
Caveat
This is an untested retirement procedure, not a measured speedup or cleanup performed today. Ronacher’s three repetitions are a smoke test, not statistical confidence. His historical hook limitations are outdated: current Claude Code documents blocking and richer lifecycle events. Do not copy the old permission-bypass or backup-deletion examples. No skill, hook, or configuration was changed.