AI Distillery
Teach AI how you think. Stop starting over.
Most people use AI the same way every time. Type a question. Get an answer. Start from zero again tomorrow. The ambitious ones write better prompts. A few set preferences. Almost nobody does what actually matters: systematically teaching the AI how they think, decide, and work.
The AI Distillery is an Agent Skill that fixes that. Feed it your raw work product (sent emails, draft revisions, feedback you've written, notes to yourself) and it distills the patterns in how you actually operate into three portable artifacts:
- User Operating Manual — an operational spec for how AI should work with you
- Preference Inference Ledger — what the AI inferred about you, with confidence levels and receipts
- Session Primer — a reusable block you paste at the start of any session to simulate memory
These artifacts are portable. They work across tools, across models, across updates. When the AI changes, your distillation doesn't.
The core principle
Don't tell AI who you are. Show it how you decide.
Self-descriptions ("I'm a direct communicator," "I prefer concise responses") are too vague to be useful and too static to be accurate. Your preferences aren't fixed traits. They're patterns that emerge from hundreds of decisions you've already made. The Distillery reads the decisions.
The engine over-infers on purpose. Your corrections are the real data.
Install
npx skills add justinneuman-coder/ai-distillery-skill
Manual install (clone + copy)
git clone https://github.com/justinneuman-coder/ai-distillery-skill.git
cp -r ai-distillery-skill/ai-distillery /path/to/your/skills/directory/
The installable skill is the ai-distillery/ folder inside this repo. Copy that folder into the location your tool reads skills from.
Before you run it
Read ai-distillery/references/SAFETY.md first. It covers data privacy (redact names before pasting), why hallucination is by design here, and a no-dump option that uses your public work instead of private material.
What it costs
Free. MIT licensed. Model-agnostic. Works in any text-based AI.
More
Built by Justin Neuman / Ultra-Normal LLC.