---
name: "learning_a_language"
description: "Coach an adult through learning a new language: input-first acquisition plan, FSRS-style spaced vocab scheduling, and low-stakes speaking from week one. Triggers on 'help me learn [language]', 'I want to learn Spanish/French/...', language study plans, or vocabulary retention problems."
---

# Learning a Language

## Purpose
Turn "I want to learn X" into a concrete, evidence-grounded study plan an adult learner can
actually sustain. The method rests on three pillars, each backed by research (see
[references/research-notes.md](references/research-notes.md)):

1. **Input-first acquisition** — Krashen's comprehensible input pitched just above the learner's
   level ("i+1"). Most study minutes go to input the learner mostly understands.
2. **Spaced vocabulary** — FSRS-style scheduling (Difficulty/Stability/Retrievability), the modern
   successor to SM-2 and the Anki default since 2023. Distributed beats massed for vocabulary
   retention across dozens of studies.
3. **Early low-stakes speaking** — Swain's output hypothesis: being pushed to produce at the edge of
   ability forces syntactic processing that input alone doesn't. From week one this means shadowing,
   self-talk, and scripted exchanges — not grammar drills.

## Workflow
1. **Intake.** Collect: target language, current level (self-assessed A0–C2 using the level ladder in
   references), daily minutes available (be honest — 20 real minutes beats 60 imagined ones), and the
   learner's *why* (travel, work, family, fun). If any is missing, ask.
2. **Level check.** If the learner is unsure of level, ask 3–5 diagnostic questions in the target
   language (e.g., introduce yourself, describe yesterday, ask for directions) and place them on the
   ladder in `references/research-notes.md`.
3. **Build the 90-day plan** from `assets/plan-template.md`. Size the three pillars to the daily
   budget using the minute-split table below. Pick concrete input sources for their language+level
   (graded readers, comprehensible-input YouTube channels, podcasts with transcripts). Verify any
   specific resource you name is real — search before naming.
4. **Set up vocabulary.** Start a deck with `bin/srs.py` (or recommend Anki with FSRS enabled for
   serious learners). Rule: add only words met *in context* (from that week's input), max
   10 new cards/day for 20-min budgets, 20/day max for 60-min budgets. Each card: target-language
   sentence on the front, answer + audio if possible.
5. **Schedule week-one speaking.** Give the learner one concrete speaking task for week 1 — e.g.,
   "record yourself answering '¿cómo estás?' in 5 sentences daily" or "do one 15-minute tutor
   session with a 10-phrase script." Keep stakes low: nobody hears it but them at first.
6. **Define the weekly review.** Every 7 days: retention check (random sample 20 due cards — target
   ≥85% recall), input upgrade (is this week's material still ~95% comprehensible? if not, drop a
   level), and one recorded 2-minute speaking sample to compare against the prior week.

### Daily minute splits
| Daily budget | Comprehensible input | Spaced vocab (reviews + new) | Speaking / output |
|---|---|---|---|
| 20 min | 10 | 7 | 3 |
| 30 min | 15 | 10 | 5 |
| 45 min | 22 | 13 | 10 |
| 60 min | 30 | 15 | 15 |

## Output Contract
Deliver, in this order:
1. A 1-paragraph read of where they are and the honest 90-day target (e.g., "solid A2: handle
   everyday exchanges, read graded material with a dictionary").
2. The weekly template (which days get which mix), filled from the minute-split table.
3. Weeks 1–2 fully specified day by day (exact materials, exact card counts, exact speaking task).
   Weeks 3–12 as milestones with level-up criteria.
4. The vocab deck state: number of cards seeded, first review schedule. (If using `bin/srs.py`,
   report the JSON path so reviews persist across sessions.)
5. One honest caveat about their specific situation (cognate distance, script barrier, speaking
   anxiety, time volatility) and how the plan compensates.

## Operating Rules
1. Never promise fluency timelines or hour counts you can't verify ("fluent in 3 months" is banned).
   Use CEFR bands and functional descriptions only.
2. Input must be *comprehensible*: ~95% known vocabulary is the working target. If the learner
   understands less than ~80% of a source, it's entertainment, not study — downgrade it.
3. Never invent resource names, app features, or study statistics. If you name a book, channel, or
   tool, verify it exists first.
4. Grammar is taught as *noticing* inside input, not as front-loaded rules. Teach a rule only when
   the learner has already tripped over the pattern twice.
5. Protect the streak: if the learner misses days, the skill shrinks the plan (drop new cards first,
   never reviews) rather than guilt-loading.
6. Speaking from week one is non-negotiable but anxiety-aware: recorded self-talk counts as speaking.
   No forced human conversation before the learner asks for it.
7. `bin/srs.py` is a teaching-grade scheduler. For real long-term study, migrate the learner to Anki
   (FSRS default since 23.10) and say so explicitly.