flashcards-word

Use when the user wants printable word / vocabulary flashcards for a young child (preschool / 5-year-old) as a Word (.docx) file — works for BOTH Chinese (hanzi / pinyin 认字卡) and English (simple words). e.g. '40 simple English words', '4 字 per page four corners', 'double-sided print', front = picture + big label, back = label only for recall/tracing. Builds cut-apart 2x2-grid flashcards with double-sided mirror alignment, emoji artwork, and a fully open-source ready, self-contained script set (Pillow + python-docx + Twemoji).

Word / Vocabulary Flashcards as print-ready Word .docx

Overview

Builds a print-ready .docx of word / vocabulary flashcards for a ~5-year-old. Works for English (simple words) and, with a pinyin-aware variant, for Chinese (hanzi). The bundled scripts/ ship the English path; the Chinese path follows the same recipe — see hanzi-flashcards-word for the pinyin-specific variant.

• N words per page arranged in a 2×2 grid (four corners) using a borderless table.
• Front page = cute illustration (emoji) + the big WORD, per card.
• Back page = the same WORD, large on white, for recall / tracing memorization.
• Double-sided (long-edge flip) mirror alignment so each back-word sits exactly under its front picture when the sheet is flipped — the part most people get wrong.
• The big word auto-scales to fit the card, so long words like "watermelon" still render without clipping.

This skill is intentionally self-contained and open-source ready: scripts/ holds a complete, dependency-only-on-common-packages pipeline (Pillow, python-docx, curl, rsvg-convert). LICENSE (MIT) and README.md are included for publishing to GitHub.

Card artwork is rendered in Pillow (PIL) (no image model needed). Word assembly is python-docx: 1 section per page, one borderless 2×2 table per page, images sized to fill each cell.

When to use

• "make 40 simple English words / vocabulary cards for a 5-year-old"
• "English flashcards, double-sided print, front word+picture, back word only"
• Any batch of English words → printable, cut-apart flashcards for a preschooler.

Don't use for: pinyin/Chinese cards (use hanzi-flashcards-word), plain vocabulary lists, adult material, or non-printable output.

Quickstart (scripts included, verified working)

bash

mkdir -p /tmp/ew && cd /tmp/ew
cp <SKILL>/scripts/words100.py .            # the word list + emoji (edit this!)
cp <SKILL>/scripts/download_emoji.py .
cp <SKILL>/scripts/generate_all.py .
cp <SKILL>/scripts/build_docx.py .
python3 words100.py                          # validate: 40 unique, multiple of 4
python3 download_emoji.py                    # emoji SVG->256px PNG (needs rsvg-convert)
python3 generate_all.py                      # render 80 card PNGs (front+back)
python3 build_docx.py                        # assemble the double-sided .docx

Env vars: EMOJI_OUT (emoji PNG dir), EN_FONT (bold word font path), OUT_DOCX.

PIL/docx work must run in a REAL Python env (a plain terminal). The execute_code sandbox usually does NOT have Pillow/python-docx — importing them there raises ModuleNotFoundError. Use the terminal tool for all card rendering, build_docx.py, and preview stitching; execute_code is only safe for dependency-free checks (regex, json, zipfile).

Environment (check once)

bash

python3 --version
python3 -c "import docx, PIL; print('ok')"   # python-docx + Pillow
which rsvg-convert    # if missing: apt-get install -y librsvg2-bin
fc-list | grep -iE "dejavu sans\b|liberation|noto sans\b"   # a bold sans for big words

Default word font: /usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf (falls back to DejaVuSans.ttf if the bold face is absent).

The word list format

words100.py exports WORDS = [(word, "", emoji_unicode_or_None), ...].

• word — the label shown large on the card (keep it short for a 5-yr-old).
• "" — the unused middle slot (kept so the tuple shape matches the hanzi version; the English renderer ignores it).
• emoji — a Twemoji codepoint to render as the illustration (can be a ZWJ/VS16 sequence; the downloader strips fe0f/200d/20e3 before the CDN filename).

Defaults ship with 40 kid-friendly words across animals / food / body / colors / nature / objects. Keep len(WORDS) a multiple of 4 for clean 2×2 pages (4 → 1 sheet, 8 → 2, 100 → 25). Keep words unique.

The 3 core recipes (details in SKILL.md / scripts)

1. Card artwork — aspect must match the on-page cell

A4 portrait, 10mm margins → usable 190×277mm; a 2×2 cell is 95×138.5mm. Per-card image aspect ≈ 95:135, so W,H = 1150,1634 (≈0.704). generate_all.py renders at 2× (2300×3268) then downscales for AA.

Each front card has 2 zones + a rounded colored frame:

• illustration (top ~35%): emoji centered, clamped inside the border.
• big WORD (lower-middle, centered): auto-fit font shrinks 24%→10% of card width until it fits within 72% of card width — long words just get smaller, never clipped.

Back card = white rounded card, one huge centered word (auto-fits to 80% width).

2. Word 2×2 grid, borderless, one table per page

• document.add_table(rows=2, cols=2), autofit=False.
• Remove all borders (w:tblBorders with w:val="none"), zero cell margins + spacing.
• run.add_picture(path, width=Mm(95), height=Mm(135)) on the cell's single run.
• Page 2+ uses doc.add_section(WD_SECTION.NEW_PAGE) — NOT a manual page-break (python-docx add_break() enum ints differ across versions and error out).
• Guard doc.paragraphs[0] — a fresh Document() may have zero paragraphs.

3. Double-sided mirror alignment (THE KEY TRICK)

Word's A4-portrait default double-sided flip is "flip on long edge" = mirror left↔right. So the back page must have its two columns swapped vs the front.

Front (TL,TR,BL,BR) = [w0, w1, w2, w3]:

text

w0  w1
w2  w3

Back (long-edge flip swaps left/right, so each back-word lands under its front-picture):

text

w1  w0     <- swap row 1
w3  w2     <- swap row 2

i.e. back layout = [(0,1,w0),(0,0,w1),(1,1,w2),(1,0,w3)].

Verify programmatically, don't trust your eyes:

python

mirror={"TL":"TR","TR":"TL","BL":"BR","BR":"BL"}
for pos,w in front.items():
    assert back[mirror[pos]]==w, f"misaligned {pos}"

Tell the user to print with 双面打印 → 长边翻转 (long-edge flip), scale 实际大小/不缩放.

Emoji artwork (fast, crisp, no art skill needed)

Download Twemoji SVG and rasterize to a high-res PNG:

bash

# per unique codepoint (strip VS16/ZWJ/combining marks from the hex filename):
curl -sL -o cp.svg https://cdn.jsdelivr.net/gh/twitter/twemoji@latest/assets/svg/<hex>.svg
rsvg-convert -w 256 -h 256 -o cp.png cp.svg      # apt-get install -y librsvg2-bin

download_emoji.py handles all of this: race-free (one temp file + dedup by unique codepoint before parallel download), md5 collision check, and writes emoji/_final.json (word → hex). Give each codepoint its own temp path — a shared _t.svg under parallel threads corrupts files.

Emoji pitfalls (all hit — must avoid)

• VS16 = 💥. Codepoints with U+FE0F (VS16) or ZWJ 404 on the CDN. Strip 0xfe0f, 0x200d, 0x20e3 before forming the hex filename; use the base codepoint.
• Shared temp file race = corruption. Don't have parallel threads write to ONE shared _t.svg; it yields identical (wrong) PNGs for different emoji. Give each codepoint its own temp path. De-dupe by UNIQUE codepoint BEFORE parallel download — many words can share one emoji (ear/hear → 👂); a late os.remove on a shared path raises FileNotFoundError. Download the unique set once, then map back every word that shares it.
• Pillow can't render color emoji from font files (0 colored pixels). Use SVG→rsvg-convert.
• Digit/keycap emoji 404 on the CDN. 1⃣–🔟 are ZWJ/VS16 sequences. For numbers, render your OWN digit icons with PIL to {EMOJI_OUT}/D{n}.png and mark the WORDS entry's emo as D1..D10 (not a real emoji), skipping the CDN pass in the downloader.

Verification checklist

• python3 words100.py → "OK: 40 unique English words" (or N).
• download_emoji.py → "resolved: N / N", no "WARN unresolved", no unintended md5 dups.
• generate_all.py → 2N PNGs (all_front/all_back), each opens clean.
• A vision_analyze sample of a few cards: each front shows the matching emoji on top + a fully-readable word (no clipping/tofu), including the LONGEST word; backs are white with one large centered word.
• build_docx.py → .docx with len(doc.tables) == 2*(N/4); mirror alignment asserted for ALL pairs (not a sample).
• Send the file (MEDIA:path) + a front/back preview image; remind: 双面/长边翻转, 实际大小, 157g+ 卡纸 for sturdiness.

Verification (structure vs content)

• Structure (deterministic, always do this): len(doc.tables) == 2*(N/4); each table rows=2, cols=2; page count = 2*(N/4) sections; total image parts ≈ 2*N.
• Content (do a vision check): sample 4–8 front cards + a couple backs; confirm emoji matches word, word fully readable (longest word in your list, e.g. "watermelon"), backs centered. python-docx renames all inserted images to media/imageN.png, so you cannot map a cell back to its source PNG by name — verify structure by count, content by eye.

Extending to a new word set

• Keep len(WORDS) a multiple of 4 and words unique. 100 → 25 sheets.
• For a FOLLOW-UP batch ("more 40, no repeats"), first recover the previous WORDS, keep them as a USED set, then assert not (set(new) & USED) AND assert len(set(new))==len(new)==N. Use a FRESH work dir (e.g. /tmp/ew2) and a distinct OUT_DOCX so files never collide with the prior batch.
• A few intentional shared illustrations across words are fine and expected (e.g. two body-part words both use ✋/👂 variants).

Publishing to the internet (open-source)

This skill is laid out to be dropped into a public repo — it works for both English and Chinese/hanzi word flashcards. The working scripts live in scripts/ (English path; the Chinese path reuses the same pipeline with pinyin in the middle slot — see hanzi-flashcards-word). The repo-root scaffolding (README, LICENSE, Makefile, requirements, env example) lives in templates/open-source/ because the skill's file-organizer only accepts scripts/ templates/ references/ assets/ subtrees — copy templates/open-source/* to the repo root when publishing, alongside SKILL.md. references/word-sets.md holds copy-paste-ready example WORDS lists (colours, animals, verbs, feelings, …) to feed straight into the README's examples or a user's words100.py.

text

flashcards-word/                          (repo root after publishing)
├── SKILL.md                            # this file (spec, recipes, pitfalls, checklists)
├── README.md                           # <- templates/open-source/README.md   (English, has examples)
├── README.zh.md                        # <- templates/open-source/README.zh.md (中文, has examples)
├── LICENSE                             # <- templates/open-source/LICENSE (MIT)
├── Makefile                            # <- templates/open-source/Makefile (make cards)
├── requirements.txt                    # <- templates/open-source/requirements.txt
├── .env.example                        # <- templates/open-source/.env.example
├── references/
│   └── word-sets.md                    # copy-paste-ready example WORDS sets
└── scripts/
    ├── words100.py          # the word list + emoji (edit me)
    ├── download_emoji.py    # Twemoji SVG->PNG downloader (race-free)
    ├── generate_all.py      # renders front+back card PNGs (auto-fit big word)
    └── build_docx.py        # double-sided 2x2 borderless .docx assembler

Publish checklist:

• scripts/*.py import words100.py from the SAME dir (via sys.path.insert(0, HERE)), no hardcoded machine paths — works from any clone location.
• Copy templates/open-source/* → repo root → git add.
• requirements.txt loosely pinned (Pillow>=9, python-docx>=0.8); README notes the system deps (rsvg-convert, a bold sans font, curl).
• SKILL.md frontmatter parseable (name: flashcards-word); description covers BOTH English and Chinese flashcard use cases.
• LICENSE (MIT) + README.md (English) + README.zh.md (中文) present at root; both have Examples / 示例 sections, cross-linked between them.
• make cards runs the 4 commands in order; optionally add a GitHub Actions workflow that builds + attaches the .docx as an artifact for folks who just want the file.
• git init → commit → push to a public repo. The .docx title is already set to "English Words · N · double-sided".

Related

• hanzi-flashcards-word — same layout pipeline for Chinese characters (pinyin).
• powerpoint — if the user later wants a .pptx deck instead of print cards.
• ocr-and-documents — text/PDF extraction if the user has source vocab to import.