---
title: Italian Fluency System
kind: side-project
role: Solo build
years: 2026 – present
status: in-progress
summary: >-
  A personal, AI-augmented system for actually speaking Italian:
  one-job-per-tool spaced repetition, an AI conversation partner, and a pipeline
  that turns native video into study decks. Design the system instead of
  grinding more hours.
tags:
  - ai
  - claude-code
  - language-learning
  - spaced-repetition
  - augmentation
order: 6
---

**The job:** Actually speak Italian, not collect another year of flashcards, on about fifteen minutes a day.

I'd done the comfortable half for years: reading, listening, vocabulary apps. Input. What I'd never really done was open my mouth, so the language sat inert. The honest diagnosis wasn't "do more hours," it was "you don't produce, and Italian doesn't live anywhere in your week." The fix is reallocation, not effort: trade one passive session for one of output, and give the language somewhere to actually show up.

So I built a system around that, and treated the design itself as the interesting part.

**One job per tool.** Recognition and production are different skills, and you get good at exactly what you practice. So Anki holds recognition only (frequency-ranked vocabulary, plus multi-word chunks), and production lives where you actually produce: cloze sentences in Clozemaster, and real conversation. No tool doing two jobs badly.

**An AI conversation partner as the speaking gym.** The reason people skip speaking is social fear, so I removed the social risk: a voice partner with infinite patience, available during a 3am feed, prompted to hold a simple conversation and correct one mistake at a time. A debrief loop then harvests the words I fumbled into my own vocabulary layer, so disposable practice becomes tracked input.

**A pipeline from native video to study decks.** I wrote a [Claude Code](https://docs.anthropic.com/en/docs/claude-code) skill that takes a native-Italian episode transcript and generates a spaced-repetition deck covering the whole episode, plus a companion list of the distinctive vocabulary worth recognizing on sight. The loop: drill the sentences (produce), harvest the vocabulary (recognize), then rewatch the episode aiming to understand all of it with no subtitles. The episode becomes the unit of study, and "I followed it without reading" is the win condition.

The whole thing is grounded in second-language-acquisition research rather than vibes: comprehensible input, transfer-appropriate processing, lexical chunks over conjugation tables, intensive versus extensive viewing. Each design choice traces back to a principle.

The principle underneath is the one I keep returning to: the leverage is in system design, in the handoff points between human and machine, not in working harder at the human part. It's the same bet as the [Agent Pipeline](/portfolio/pipeline-skills) and [Work Context Protocol](/portfolio/work-context-protocol), pointed at a hard human skill instead of at code. See [/writing](/writing) for the augmentation thinking behind it.