Plan Words Rules Sentences

Plan: a phrasebook, then words, then rules, then sentences

Source of truth: this wiki page.

Written 2026-10-07. With the B1 plans (b1-plans.md), owner's order: paths are rewritten then anyway. No interim change to lessons before it (owner's choice).

What the owner asked

Check the current deck pacing of telugu. It feels i am learning way too many telugu sentences even before having a decent vocabulary. Is there any research on the learning levels before moving on to the next stage?

Then:

I think ideally we should not teach sentences beyond ultra basic phrases (greetings and i don't know telugu types) before teaching all their words and grammar should start with word forms, conjunction, preposition etc rules. The current preposition teaching was just 3, and always with the same words, not in a rule form. I have completed 6 decks now. And feel lost instead of feeling like i made progress.

Decided with the owner:

• A small phrasebook is taught whole from the first lesson: 15–25 survival chunks per language.
• Any other sentence waits until every word in it, by its base word, and every rule it uses are known.
• Known means mastered: the learner's last review of a word or rule, in each skill asked, was right and FSRS gave at least 85% recall at that review, the same known as Progress (owner, 2026-10-24: "Align 'known' with FSRS"; it replaced 9 of the last 10 answers right, 2026-10-16).
• Grammar is taught as rules, starting with word forms, endings and postpositions, and conjunctions, practised across the words the learner knows.
• Grammar decks are about the rules, not the words: "Grammar decks should focus on the rules rather than the words."
• A grammar question chooses among forms of the same word: "The grammar drills should be like choosing between ammato, ammaki, ammalo, etc. Not ammato and paalu."
• When: with the B1 plans.

What exists

Measured on Telugu (main, 2026-10-07):

• The first eight units, before the script, hold 95 single words, 47 sentences and phrases, and 77 grammar cells. Unit 1 is 6 words and 16 sentences.
• Every lesson takes three "hard" items in nine (lessonItems, ADR-0024): phrases, sentences of three words or more, and grammar cells. The pool is the next two unfinished units, interleaved, so sentences come from the first lesson on.
• The sentences' words are mostly untaught: 0% known in every sentence of the first six lessons; 0–50% by lesson 12, counting exact word forms.
• Grammar decks are fixed phrases, not rules:
  • "Endings for in, to and with" is three phrases: గదిలో (gadilō, in the room), బడికి (baḍiki, to school), అమ్మతో (ammatō, with mother);
  • the postpositions deck is 10 fixed phrases, never applied to other nouns. The rule is only in the notes.
• What the owner had finished: six decks, about 33 single words and 32 sentences or phrases.

The research

No study gives a fixed number of words to learn before sentences, but the evidence points one way. It was checked from abstracts and search records, not full texts. Confidence: High means the citation and finding were both seen in an abstract or repository record; Medium means the finding came from a summary.

Finding	Source	Confidence
Understanding a text needs 95–98% of its words known. 80% was the lowest coverage tested	Hu & Nation 2000; Laufer & Ravenhorst-Kalovski 2010; Schmitt, Jiang & Grabe 2011 (no sharp threshold: comprehension rises with coverage)	High
Listening needs about 95%	van Zeeland & Schmitt 2013	Medium
Guessing a word from context needs about 95% of the others known	Liu & Nation 1985	Medium
The more unknown words in a sentence, from 0 to 7, the worse it is understood, especially beyond 5	de la Garza & Harris 2017	High
For new words, a translation, alone or in one short sentence, works as well as or better than a longer context	Prince 1996; Laufer & Shmueli 1997; Webb 2007; Webb, Yanagisawa & Uchihara 2020 (meta-analysis)	High
Recalling a meaning beats inferring it from context, for retention	van den Broek et al. 2018; Mondria & Wit-de Boer 1991	High / Medium
Fixed chunks help from the start, and are later broken into parts and reused	Wray 2002; Myles, Hooper & Mitchell 1998	High
A first "survival" syllabus of about 120 words and phrases	Nation & Crabbe 1991	High
Mastery learning (80–90% before moving on) raises scores by about half a standard deviation, at a cost in time; the effect on standardised tests is weaker	Bloom 1968; Kulik, Kulik & Bangert-Drowns 1990; Slavin 1987	High
A grammar structure can only be taught once the learner is ready for it. No sequence exists for Telugu	Pienemann 1984, 1998; Spada & Lightbown 1999	High / Medium
Deliberate study should be about a quarter of a course; input and fluency work need almost every word known	Nation 2007 (Four Strands)	High
The CEFR gives no word counts; for English, A1 is under 1,500 words and B1 is 2,750–3,250	Council of Europe 2001; Milton & Alexiou 2009	Medium

What follows for this plan (inferences):

• A sentence whose words are all known is what the evidence supports. The owner's rule, that every word and rule is taught first, meets it. In a sentence of four to six words, a single unknown word already means only 80–83% known.
• Phrasebook chunks need not wait for their parts.
• Words are best taught first as words, with their translations.
• "Known" should be counted by stem and ending, not by whole word: a Telugu word can hide several new parts. The coverage figures were measured on English and are untested for Telugu.

The design

1. A phrasebook

• A card marked phrasebook: true is taught whole and is never held back:
  • greetings, thanks, sorry;
  • yes and no;
  • "I don't know Telugu", "I don't understand", "please speak slowly";
  • and the like.
• 15–25 per language, taught in the first lessons.
• Its words are not counted as taught by it: each word comes again as a word, and its base words (b1-plans.md) say which.

2. Words

• Every word is taught as a word before any sentence that uses it, with its base form, reading and meaning. A unit's words come first in that unit.

3. Rules

• A rule card is a new kind of item. It holds:

  • the rule's name;
  • its explanation, with readings, e.g. "-లో (-lō) means in; it joins the noun's oblique stem: ఇల్లు (illu) → ఇంట్లో (iṇṭlō)";
  • the forms it makes.

• Taught, then practised as a table over known words:

  • the rows are the words of the right kind the learner has been taught (nouns for endings, verbs for person endings);
  • the columns are the rule's forms (-లో, -కి, -తో, -నుంచి);
  • each cell is a grammar question.

  So a postposition is practised on every known noun, not on three fixed phrases.

• Its questions test the rule, never the word:

  • Choosing: the options are the forms of the same word, అమ్మతో (ammatō), అమ్మకి (ammaki), అమ్మలో (ammalō), అమ్మ నుంచి (amma nuñci), never a different word such as పాలు (pālu). The meaning to express is the prompt: "with mother".
  • Typing: the word and the meaning are given; the learner types the form.
  • The word itself is always one the learner knows, so only the rule is being tested. The grammar-table format already expands a table into cards; what is new is the rule card, and rows filled from known words.

• First come word forms:

  • nouns: plural, the oblique stem, the case endings and postpositions;
  • pronouns and their forms;
  • verbs: person and tense endings;
  • then conjunctions.

4. Sentences

• A sentence names the rules it uses (rules: [te-rule-lo, …]), beside its base words (bases, b1-plans.md).
• It unlocks once all of them are known, at the mastery bar. Until then it is not offered, and does not hold back its unit.
• Reviews never wait: the bar decides only when new sentences appear.
• Sentences are built from what is known, so most are practice, not new material.

5. Lessons

• The fixed three of each difficulty goes.
• A lesson's new items are the next eligible ones in the path's order:
  • phrasebook;
  • the unit's words;
  • the rules whose words are taught;
  • the sentences now unlocked.
• Progress shows in those terms: words, rules and sentences unlocked, so that a learner sees what they can now build.

6. Paths

Each unit of a B1 plan is its theme's words, then the rules they need, then its sentences. Each rule counts once toward the plan's grammar topics.

7. Existing decks

Converted as each language's B1 plan is written:

• first-words splits into the phrasebook and words;
• each grammar deck becomes rule cards and tables;
• its sentences are tagged with their rules and bases, and move behind them.

What it takes

1. Format:
   • the phrasebook mark;
   • rule cards;
   • rules on sentences;
   • tables whose rows come from known words;
   • the parser, the validator, docs/DECK-FORMAT.md, and an ADR.
2. The validator:
   • every rule a sentence names exists;
   • every phrasebook list is 15–25 cards;
   • a sentence's words resolve to cards or bases (b1-plans.md).
3. Lessons:
   • no fixed quota;
   • the unlock gate, with the mastery bar per word and rule;
   • rule tables filled from taught words;
   • a grammar question's options are the other forms of the same word.
4. The rule card's drill: taught, then its table's cells as grammar questions.
5. Content, per language with its B1 plan:
   • the phrasebook;
   • the word-form rules;
   • sentences tagged.
6. Tests:
   • a sentence held back until its last word and rule are taught;
   • the phrasebook never held back;
   • a rule's table growing as nouns are learned;
   • a grammar question offering only forms of its own word;
   • lessons without the quota.

Decided, 2026-10-16 (#414)

• The mastery bar: "Make that 90% of 10". A word or rule is known when 9 of its last 10 answers are right. Replaced 2026-10-24 ("Align 'known' with FSRS", #416-#418): known is the Progress star's FSRS known, the last review of each skill right with at least 85% recall at that review (kKnownRecall, leavesKnown). A rule is known when every skill asked in any cell of its table is.
• How a rule makes its forms: "Stem + exceptions, checked", replacing "every form listed per word" (2026-10-09; b1-plans.md).
• Cells of a growing table per lesson: "3 per lesson".
• When: still with the B1 plans ("Keep with B1 plans"), after the owner said "We need to fix the paths" and "First two weeks" for when Telugu felt lost.

Built

• Lessons without the quota, and the unlock gate (#416, ADR-0046): the next eligible items in path order, fewer while words and rules in progress are below the bar; a sentence waits for its words and rules; a rules table's cell is taught only for a known word; the unit page greys a closed sentence and a tap opens "Why it is closed". Owner, 2026-10-23: "Greyed + why sheet; no plan", so a lesson starts straight away with no plan screen.
• The rule card and its drill (#417): a rule is taught on its own card (name, explanation, its forms on known words) before the first question about its cells; a grammar question is chosen first, then typed, among forms of the same word; after an answer the word's forms in the rule are shown with the cell asked marked; a rule's table on the unit page shows only known or asked words, each cell known, learning or not asked yet.
• Progress in words, rules and sentences (#418): the Progress tab's three tiles in place of Words, counted against what is taught so far with words by base word, then the Level tile alone, one slim row, the week and the star, with the hours tiles under the star so that it stays on the screen at 1.4x text; "Unlocked in this lesson" at the end of a lesson, counts only (owner, 2026-10-23: "the lesson end shows counts"); and the unit page's counts (#416).

Estimate

Part	Hours
Format, parser, validator, ADR	4–6
Lessons: gate, no quota, rule tables	4–6
Rule card drill	2–3
Content per language: phrasebook, rules, tagging	4–8 each

Confidence: medium for the app; low for content.

Script and word core, from the Anki comparison (owner, 2026-10-17)

Asked where the script goes in every path: "After units 4-6". The script decks follow the phrasebook and the first word units, within the first two weeks; a learner who already reads the script skips them (#554).

Asked whether a frequency-ordered word core comes before the big grammar tables: "Yes, smaller (~200)". About 200 common words, ordered by frequency (the Wikipedia counts, #516), in themed units before the large tables, so the tables use words the learner knows.

Every path is rebuilt to this design: phrasebook, the first word units, the script, the ~200-word core, then rules as forms and set cards, with sentences waiting until their words and rules are known (owner, 2026-10-17: "build the path files for all present languages via an opus agent, as per our current path design"). Now Haiku writes decks and Sonnet reviews them (Opus only if the owner asks, 2026-10-22).

See also: Roadmap · Plans-Index · Home