Preference Sync

Most corrections you give Claude — too long, too formal, wrong format, don't say "delve" — are the same three to five corrections you gave it yesterday and will give it again tomorrow. Each correction takes 30 seconds and the cumulative cost over a year is substantial. This skill captures them once, encodes them into a preference file, and stops the loop.

The extraction runs in three phases. Explicit preferences first — direct questions on voice, format, length, banned phrases, preferred phrases. Implicit preferences next, mined from past corrections — "what's the last thing Claude did that annoyed you?" and "what do you typically edit out before using the output?" The implicit pass is where the real material is. People often can't articulate their preferences until they're shown what they keep cutting. Finally a calibration test — two contrasting sample outputs generated against the preferences gathered so far, with iteration until the voice locks. The calibration step is non-negotiable. Shipping preferences without testing them on a real output is how you end up with a file that doesn't actually change behaviour.

Negative preferences carry more weight than positive ones. What you don't want is often more defining than what you do. "Banned: 'it's important to note', three-part vague-noun lists, any sentence that starts with 'In today's'" is more actionable than "be direct." Specific anti-patterns are testable; aspirational rules ("write well") are not.

The output is a structured preference file — Voice Profile, Format Rules, Content Rules, Domain-Specific Overrides, Anti-Patterns — designed to live in CLAUDE.md, in a memory file, or in a dedicated preferences document. Overrides matter. Some contexts need different rules: a client project might require a more formal tone than your default, a fundraising email needs different defaults than a Slack message. The file supports project-level and task-level exceptions rather than forcing one voice for everything.

Preferences evolve. The file has a version number and an update date because what works today may not work in six months. The skill also infers from corrections mid-session — if you edit Claude's output two or three times on the same dimension, it proactively suggests adding the rule to the preference file rather than leaving you to repeat the correction next session. The end state is a file that gets revised quarterly, not one that gets written once and abandoned.