The LLM-to-rules ratchet: distilling agent-discovered patterns into deterministic sensors

## Spark

From the 2026-06-27 digest. Meta's asset-classification write-up names the exact pattern I keep reaching for: use the LLM to interpret novelty, then distill what it learns into versioned, auditable rules that run *without* the model — steadily shrinking the LLM's role over time.

## Why it could be interesting

The activity is "I write Claude skills and hooks." The claim is sharper: an LLM in a durable system should be a *ratchet, not a runtime* — every stable pattern it discovers gets promoted to a deterministic sensor (a hook, a lint rule, a committed CLAUDE.md memory, a skill's scripted fan-out), so the model's surface area shrinks as the system matures, rather than the system depending on the model forever. That's the opposite of "wrap everything in an agent."

## Open questions

- [ ] Concrete worked example from my own repos: a pattern that started as agent judgement and became a committed rule (Vale accept.txt? a pre-commit hook? a digest filter?).
- [ ] Where the ratchet *shouldn't* click — which judgements stay with the model permanently, and why.
- [ ] Is this distinct enough from #67 (skills are pipelines) and #189 (sensors), or a synthesis that supersedes them?

## Source material

- "When the system finds stable, validated patterns, those should become versioned, auditable rules that run without the LLM" —
- "distill into versioned human-reviewed deterministic rules, which steadily shrinks the LLM's role in production" —
- "automations are an accessibility tool" —
- Own evidence: blog #268, chezmoi #371 (AI-docs positive-framing pass)
- Related: #67, #189, #96