Ten folders · one push · 30–60 minutes
A training desk, not a curriculum
Your library already contains two methods that never agreed with each other. This page settles the argument the way language schools that answer for their results settle it — and then gives you something to actually do today.
01 — Reconciliation
The two day-plans in your repo, merged
METHOD.md gives a day that is 50% recall on your own sentences, 30% grammar book, 20% input. notes/how-to-learn-languages.md gives a different day: grammar spine, captioned video, a frequency deck, two tutor sessions a week, minimal-pair drills. Each file calls itself “the day.” Neither contains the other’s largest component — there is no SRS line in the first and no sentence-recall line in the second. Across the two files and the README there are four different figures for how long a day takes: 15 minutes, 30 minutes, 45–60 minutes, and a routine table totalling up to three hours.
The military schools resolved this a long time ago, and not by choosing. A day at the Defense Language Institute is roughly six classroom hours that move through several different activities with different purposes — and the thing that makes it one day rather than a pile of tactics is that each block ends in something you could not do before. Your two files are not rival philosophies. One is a production engine and one is a comprehension engine, and you need both, in that ratio, in that order.
| Block | Min | What | Which engine |
|---|---|---|---|
| Cold open | 10 | Grammar spine, every exercise attempted closed-book before checking | Comprehension — the largest single effect in the literature |
| Drill floor | 15 | Active recall on your own sentences, out loud, from English | Production — METHOD.md’s engine, unchanged |
| Dead time | — | Sentence audio, then graded listening. Costs no scheduled time | Comprehension — free hours |
| Evening | 10 | Frequency deck, L2→L1 until the spine is done, then flipped | Comprehension — understanding other people |
| Twice weekly | 30 | Tutor, with the standing correction instruction | Both — the only block that supplies feedback |
| First fortnight | 10 | Minimal-pair perception, many different speakers | Perception — then it stops and never returns |
notes/how-to-learn-languages.md overturns METHOD.md on pronunciation and
prescribes generating sentence audio in several different voices, because the number
of distinct talkers is what makes perception training work. But
build_audio.py takes a single --voice, and its cache key is built
from that one voice string, so multi-voice output cannot currently be expressed. The
correction is written down and unimplemented.
recall_drill.py caps its Leitner ladder at 14 days, so a fully-mature sentence
returns every fortnight forever. The SRS arithmetic in your own notes assumes mature cards
sit at 60–150 days. At a 14-day ceiling a thousand-sentence list costs about 71 reviews
a day indefinitely — roughly five times what you budgeted. The drill below uses the
longer ladder.
02 — Drill floor
Say it out loud. Thinking it does not count.
All 32 sentences from Sentence Islands/Spanish/sentences.csv, live. Progress
saves in this browser. Four modes, each borrowed from a different part of the schoolhouse
day — recall is your own method, the other three are the ones nobody self-studying
ever trains.
Ready
Press start.
Space reveals the answer. 1 missed it, 2 had it, R replays audio.
Recall
The engine. Getting it wrong and then seeing the answer is what builds the memory — a re-read of the answer you never tried to produce does almost nothing.
Shadow
Speak simultaneously with the audio, not after it. Builds mouth timing at full speed. The thinnest evidence base of anything here — no meta-analysis exists.
Gist
One pass, no replay, then say what it was about in English. This is the actual intelligence-linguist skill, and it is the one self-studiers never practise.
03 — Technique library
What transfers to thirty minutes alone
Two facts frame everything below. The first is that the Foreign Service Institute measured this in 1967 and found a student working alone with recorded material participates with the language at 400–600 times an hour, against 100–150 in a class shared with three other people. For the drilling layer, working alone is denser, not degraded.
The second is that the Army's own sustainment requirement for a qualified linguist is 8–16 hours a month plus one annual block of 160 consecutive hours, against a written individual plan, with a real test at the end. That is the military's own answer to “what keeps a professional at level,” and unlike sixty weeks of full-time school, it is a dose you can actually hit.
One caution before you start collecting. The drills people associate with the “military method” are a 1940s inheritance the institution has explicitly moved past — its current catalogue describes task-based, content-based, learner-centered instruction, and pattern drills appear nowhere in it. Its instructor certification course does not teach them. What is worth taking is the architecture — short cycles that end in real use, authentic material from day one, diagnose-then-target — at least as much as the drills.
A week of it
Scaled to 30 minutes a day, which lands inside that 8–16 hours a month.
| Day | The block |
|---|---|
| Mon | Microwave cycle on new material |
| Tue | Shadowing, then minimal pairs |
| Wed | Dictation or transcription, then the named-entity log |
| Thu | Scan-and-search, or one complete GLOSS lesson |
| Fri | Transformation drill, then self-run question and answer |
| Sat | Two-pass listening, then sight translation |
| Sun | Interview self-simulation monthly; otherwise domain reading |
| Daily | Additive, not a substitute: 10 minutes of spaced retrieval |
| Quarterly | Re-diagnose, and take one full iso-immersion day |
| Yearly | Your own 160-hour block: one full week of concentrated study before a self-test |
04 — The ladder
Where the honest benchmark actually sits
The Interagency Language Roundtable scale is the one the US government actually pays against, and it is worth internalising because it is far less flattering than the scales used in language marketing. Level 2 is called limited working proficiency. It is where most graduates of a full-time, sixty-week, selectively-admitted military course land.
Fixed phrases. Cannot recombine. Most app streaks live here.
Survives predictable everyday exchanges. Frequent errors, constant repetition needed.
Handles routine work demands and concrete topics. Still breaks down on abstraction, argument and unfamiliar register. This is the military graduation target — and roughly 3 in 10 people who start the course do not reach it.
Only about a quarter to a third of graduating classes reach this in listening and reading together.
Abstract topics, argument, unfamiliar subject matter, structural accuracy that never impedes. Government auditors put linguist productivity at up to four times higher here than at level 2. This is what people mean when they say “fluent.”
Level 5 is an articulate educated native speaker. It is not a study goal.
The dosage, honestly
A Category IV course — Arabic, Chinese, Japanese, Korean — is 64 weeks at six classroom hours a day: roughly 1,920 contact hours, plus homework, with students pre-selected by aptitude test and studying as their full-time job.
At 30 minutes a day, 365 days a year, you accumulate 182 hours a year. The same hour count would take about ten and a half years — and that assumes perfect linearity, which nobody believes.
What that buys, and what it doesn't
Those six hours a day are not mostly buying vocabulary. They buy processing automaticity under time pressure and thousands of hours of interactive output with someone correcting it. That is the part 30 minutes a day cannot replicate.
What does transfer — and where the distributed learner genuinely has the advantage — is spacing. In a nine-year study, 13 sessions at 56-day intervals matched the retention of 26 sessions at 14-day intervals. Half the work, same result, for the memorisation layer specifically.
| Cat | Your folders | Class hrs | Still quoted | Note |
|---|---|---|---|---|
| I | Spanish, French, Portuguese | 552–690 | 600–750 | Closest to English. Spanish then Portuguese then French compounds the discount. |
| II | — | 828 | 900 | German, Indonesian, Swahili. None of your ten. |
| III | Russian, Hebrew, Farsi, Tagalog | 1,012 | 1,100 | Hard grammar or a new script, but no character set. About forty languages sit here. |
| IV | Arabic, Chinese, Japanese | 2,024 | 2,200 | Script and distance. Arabic is really two projects — a formal register plus a spoken dialect. |
Use all of these as an ordinal ranking and nothing more — which is the
conclusion notes/how-to-learn-languages.md already reached independently. They are
class hours only; the same courses expect a further 17 hours a week of self-study, and they
target level 3, a full step above what the military course aims at.
05 — Armory
Free, official, and better than most paid material
All of it is US government work, all of it is free, and most of it is built for exactly the skills in the library above. The reading and listening lessons are graded by the same ILR levels on the ladder, which means you can aim at a rung rather than at a vibe.
06 — Ship the page
Four weeks to this page on a public URL
The full walkthrough is in notes/how-to-learn-github-and-the-web.md next to this
file. This is the checklist. It ticks and saves. One rule: commit every day you work, even
if it is one line — you cannot learn version control by reading about version control.