Second Brain

setup-curator skill + triad vs passive skill-only approach

2 min read

setup-curator skill + triad vs passive skill-only approach

Decision

Build setup-curator as an enforced skill with a 3-layer automation triad (Hookify autoloader + PreToolUse write-guard + nightly Pristine Sweep cron) rather than a passive skill referenced in CLAUDE.md.

Rationale

A passive skill relies on model discipline every session. Over dozens of sessions, drift accumulates — someone edits MEMORY.md without routing through the skill, counts go stale, duplicate content creeps in, truncation cliff approaches silently. The compounding-intelligence aspiration (“self-organizing, self-correcting, self-healing, permanently pristine”) requires enforcement that operates independently of the current session’s discipline. The triad achieves this with zero LLM cost at enforcement time (all three layers are deterministic Python/regex) and zero false-positive rate (advisory warnings, not hard blocks, so trivial edits proceed). First-run detection of genuine drift (4 findings on first sweep, all fixed) validated the approach works.

Alternatives Rejected

Option A (rejected): Skill-only — Add setup-curator as a regular skill, reference in CLAUDE.md, hope model invokes it. Risk: drift accumulates between sessions with no detection.

Option B (rejected): Hard-blocking guard — Block all Write/Edit on setup files unless skill invoked. Risk: annoyance on trivial typo fixes, friction for legitimate quick edits.

Option C (chosen): 3-layer triad with advisory guard + nightly sweep — Auto-loader surfaces skill proactively, guard warns without blocking, cron catches accumulated drift. Balances enforcement with usability.

Outcome

Pending

Graph View

Backlinks