French Sentence Breakdown

French Sentence Breakdown & Study System

A skill for converting raw French audio/text into structured bilingual study material — using English as the bridge language, not the learner's native language — and for running an ongoing French-acquisition system built on the Three Listen Rule.

Why English as the bridge, not the learner's L1: French and English are both Indo-European and share grammatical categories (tense marking, agreement, relative pronouns, article systems) that a non-Indo-European L1 often lacks. Translating French into English tends to preserve the grammatical structure in a way translating into, say, Chinese does not — the L1 translation is often correct in meaning but strips the grammar out entirely. Use this only when the learner is genuinely comfortable in English at a level where it can serve as a reliable bridge — confirm this once rather than assuming it.

This convenience comes with three real risks, so every breakdown must actively guard against them (see Step 2):

1. Translation-mediated reaction speed — routing French through English on every sentence can

become a crutch that slows real-time listening/speaking, where there's no time for a two-step conversion. Counter with a mandatory French-only restatement step.

1. Faux amis (false friends) — French and English look alike, which makes look-alike-but-wrong

word pairs the single biggest failure mode of this method (actuellement ≠ actually, assister à ≠ assist, librairie ≠ library). Counter with a running false-friends log.

1. Structure calquing — copying English sentence structure into French where the two actually

diverge (e.g. plaire à qqn reverses subject/object versus "to like"; passive constructions like "I was rejected for a visa" need restructuring in French, not word-for-word translation). Counter by explicitly flagging divergence points, not just convergence points.

This skill has two modes. Detect which one the input calls for:

1. Material mode — the user gives you French content (transcript, article, lyrics, their own

journal entry) and wants it cleaned up, explained, corrected, or turned into study material.

1. System mode — the user wants a listening/study plan, a progress check, or help deciding what

to practice next.

Most conversations start in Material mode and drift into System mode once several pieces of material exist. Keep both available.

---

Mode 1: Material Mode

Step 0 — Classify the input

Before doing anything, identify:

• Is it clean or raw? Auto-generated transcripts are full of noise: missing punctuation, wrong

homophones, dropped accents, stray [musique]/sound-effect tags, English asides, mis-heard proper nouns. If raw, clean it first (see references/transcript-cleaning.md).

• Is it teaching material or authentic material? A didactic podcast (slow, repetitive, built

for learners) gets treated very differently from a real vlog/interview/song (natural speed, elision, disfluencies). Say explicitly which type it is and roughly what CEFR level it sits at (A1–C1) — the user needs this to calibrate expectations. Don't let them assume every piece of content is equally learnable at their current level.

• Is it the user's own writing? If so, this is a correction task, not a transcript-cleaning

task — go to references/writing-correction.md instead of the breakdown format below.

Step 1 — Produce a clean, speaker-labeled, thematically segmented transcript

Only needed for raw/messy input. Standard for a clean transcript:

• Segment into numbered thematic sections with a short French header (e.g. ## 3. Le dessert)
• Label speaker turns if multiple speakers (infer from content if the raw file has no labels —

say so explicitly, e.g. "attribution reconstructed from content")

• Fix mis-transcribed words, missing accents, dropped partitive articles, etc.
• Append a corrections table at the end: | Sous-titre automatique | Correction | Type |
• If anything is genuinely inaudible or a musical/sound-effect artifact, cut it and note it in the

table rather than guessing at content

Full template and worked example: references/transcript-cleaning.md

Step 2 — Sentence-by-sentence breakdown

This is the core deliverable. For each sentence, produce the six-part format below. Full template, worked examples, and calibration notes (when to compress a short/simple sentence) are in references/sentence-breakdown-format.md. Read that file before producing breakdowns — it defines exactly how much detail each part needs and when to abbreviate.

The six parts, in order, per sentence:

1. English translation
2. IPA
3. Word-by-word gloss (number every word/morpheme)
4. Sentence structure (skeleton, for anything beyond a simple SVO sentence)
5. Grammar notes (the grammar point(s) worth teaching from this sentence — the actual payload).

Explicitly mark each point as either converges with English (structure maps over cleanly — worth noting briefly, since this is exactly why the bridge works) or diverges from English (worth slowing down on — this is where calquing errors come from). Don't only list convergences; divergences are the higher-value half of this section.

1. Extension examples (2 short examples using the same structure)

Two additions run alongside the six parts, not per-sentence but as running threads through the whole breakdown:

• French-only restatement. After finishing a natural chunk of sentences (a themed section, or

every 5-8 sentences in continuous prose), insert a short prompt telling the learner to close the English side and restate the passage in French only, in their own words. This is not optional decoration — it's the step that converts the English-bridge comprehension gain back into direct French processing, and skipping it is exactly how the bridge becomes a permanent crutch.

• Faux-amis log. Maintain a running table across the whole document of any French word that

resembles an English word but means something different, or where the surface similarity could tempt a wrong guess. Append to it as they come up rather than scattering the warning inline only — a consolidated list is what the learner will actually review later.

Do not give every sentence the full six-part treatment. Short/simple/filler sentences (single interjections, exact repeats of a structure just covered, "Incroyable.", "Trop bon.") should be compressed to 1–2 lines. Spend the words on sentences that teach something. See the compression rules in references/sentence-breakdown-format.md.

Handle long transcripts in one pass. The user has explicitly asked for longer sessions rather than many tiny ones — a 40–80 sentence transcript should normally be broken down as one continuous document (or at most 2–3 large parts if length limits force a split), not chopped into many short files. Don't default to 15–20 sentence chunks unless the user asks for that.

Step 3 — Cross-cutting grammar review

At the end of each breakdown, add a grammar review table: every grammar point taught, which sentence(s) it came from, whether it converges or diverges from English, and a one-line takeaway. See the format and the "recurring structures" style analysis in references/sentence-breakdown-format.md — for longer/authentic texts, explicitly call out which structures recur across many sentences (these are the highest ROI to master) versus which appear once (lower priority, recognition-only for now). Close with the accumulated faux-amis log table so it reads as one reference list, not scattered footnotes.

Step 4 — Practical use note

Close with a short, concrete note on how this specific piece of material should be used given its difficulty level relative to the learner (intensive deep-dive vs. extensive/background listening vs. skip-for-now). Don't let a B2 vlog get treated the same way as an A1 teaching podcast — say so.

---

Mode 2: System Mode

When the user wants a study plan, a progress check, or help deciding what to do next, use references/study-system.md. It defines:

• The Three Listen Rule, scaled correctly by learner level and by material type (teaching

podcast vs. authentic content) — this is a listening comprehension technique and must not be conflated with extensive/volume listening

• A daily engine (vocabulary SRS, micro-listening, self-breakdown, extensive listening, output)

that fits into a fixed daily time budget

• A back-translation method (French → English → French, spaced across days) for converting

listening material directly into writing practice, plus the French-only restatement drill that keeps the English bridge from becoming a permanent crutch

• A systematic grammar curriculum (ordered list of grammar points to cover on a timeline) to

run alongside the material-driven, extraction-based grammar notes from Mode 1

• Level-appropriate material tiers (staple / bridge / stretch) and the upgrade test for moving

between them

• A weekly rhythm and a monthly checkpoint with concrete pass/fail thresholds

Always ask or infer: current vocabulary size, current CEFR level, target level, deadline, daily time budget. These five numbers drive every recommendation — don't give generic advice without them. If any are missing, ask once, briefly, rather than guessing.

---

Formatting conventions (apply in both modes)

• All learner-facing explanatory text is in English; French sentences, IPA, and example

sentences stay in French. (English was chosen deliberately as the bridge language — see the rationale above — not defaulted to as a generic choice. If the learner's preference changes, or if a specific phrase is genuinely clearer glossed in their L1, that's fine as an occasional aside, but it shouldn't become the default mode.)

• Bold the French target word/phrase inside English explanations rather than translating it away.
• Use > blockquotes for important warnings/exceptions the learner is likely to get wrong —

divergence points and faux amis are the most common candidates for this treatment.

• Never present output only as a chat reply for anything long-form — create a markdown file so the

user can save/reuse it (this is study material, not a one-off answer).

• Respect copyright: this skill produces learner analysis and pedagogical restructuring of the

source, not a verbatim reproduction for its own sake. Never reproduce song lyrics. Transcript cleaning is fine because it's a functional transformation (fixing ASR errors) in service of language instruction, not redistribution of the work.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Neofirst0715
• Source: Neofirst0715/french-sentence-breakdown
• License: Apache-2.0
• Homepage: https://neofirst0715.github.io/french-sentence-breakdown/

Install and usage instructions live in the source repository linked above.