In The Cards Lenormand Skill

Interpret Lenormand card readings from freeform spreads or compatible Stargazer’s Oracle prompts, persist each interpretation as structured JSON plus human-readable Markdown, and generate polished A4 PDF, web, social-card, or long-image products. Uses a three-phase workflow: parse → interpret and save → generate. Supports six spread types, combination-first reading, bilingual Chinese/English output, natural-language retrieval, and regeneration of past readings. Triggers on: "牌间 / In the Cards / Stargazer 解读 / 帮我解读这个牌阵 / 雷诺曼解读 / Lenormand reading / Stargazer prompt / 帮我解牌 / 帮我解读 / interpret this spread / 生成解读 PDF / 列出最近的解读 / 把上次的解读 / 生成网页版 / 生成社交卡片 / 上次的解读 / 之前的解读 / 請為我解讀 / 幫我解讀 / 每日運勢 / 今日運勢 / 我想透過 Lenormand / 透過 Lenormand / 二選一 / Lenormand 卡牌占卜 / consult the Lenormand cards / compare them with Lenormand / daily fortune / read my Lenormand / deciding between two options".

Linda Huang@locoda

Install

openclaw skills install @locoda/in-the-cards-lenormand

In the Cards · 牌间

Lenormand Interpreter

Interpret Lenormand card spreads, persist them as reusable data, and generate visual products — A4 PDF, web pages, social cards, or long images — by copying seed HTML files and filling in placeholders. Accepts freeform spreads and prompts from Stargazer’s Oracle.

Architecture

text

Lenormand Prompt
(Stargazer-compatible)
    │
    ▼
[Phase 1] Parse → structured card data
    │
    ▼
[Phase 2] Interpret → save reading (JSON + MD) to output/readings/
    │
    ▼
[Phase 3] Generate products from saved reading
           ├── A4 PDF (default, Kami style)
           ├── Web Page (standalone interactive HTML)
           ├── Social Card (1080×1440 carousel, 3:4)
           └── Long Image (full-page PNG screenshot of web page)

Every reading is saved before product generation. Saved readings can be regenerated into any product type later.

When to Use

First-time reading: User pastes a Lenormand spread or compatible Stargazer prompt and asks for interpretation. → Run Phase 1 → Phase 2 → Phase 3 (generate default A4 PDF).

Product regeneration (post hoc): User references a past reading and asks for a different product type. E.g., "把上次的解读生成社交卡片" / "用 #3 解读生成网页版" / "列出最近的解读". → Skip Phase 1–2, load the saved reading JSON, run Phase 3 with the requested product type.

Phase 1: Parse the Prompt

AGENT — REQUIRED BEFORE PARSING: Execute these two file reads now, in order, before doing anything else:

Open and read references/spread-parsing.md in full

Open and read references/lenormand-cards.md in full

Do NOT begin parsing the prompt until both files have been read and their contents are in your context. Using training knowledge as a substitute is not permitted.

Read references/spread-parsing.md for detailed parsing rules. Extract structured data:

Strip universal instruction prefix — the "You are a traditional Lenormand reader..." / "你是一位傳統 Lenormand 占卜師..." block before ---
Detect language (ZH/EN) from 【問題】 / Question: / Spread:; fallback: CJK character ratio.
Detect spread type: yesno-1 | two | three | five | nine | choice
Extract fields: question, date, cards with positions, verdict (yes/no), option A/B texts (choice)
Resolve card names against references/lenormand-cards.md
Output a structured data object

AGENT — PHASE 1 CHECKPOINT: Before proceeding to Phase 2, emit the following block exactly as shown, filled with the parsed values. Do not continue until this block is visible in your output.

text

PHASE 1 COMPLETE
─────────────────────────────
spread_type  : {yesno-1|two|three|five|nine|choice}
language     : {zh|en}
question     : {extracted question text, or "daily fortune" if null}
date         : {YYYY-MM-DD}
cards        :
  {position_label}: {card_number}. {card_name_zh} / {card_name_en}
  ... (one line per card)
verdict      : {YES/NO/MAYBE, or N/A}
option_a     : {text, or N/A}
option_b     : {text, or N/A}
─────────────────────────────
Proceeding to Phase 2.

Phase 2: Interpret & Save

AGENT — REQUIRED BEFORE INTERPRETING: If references/lenormand-cards.md is not already in your current context from Phase 1, open and read it now before writing any interpretation. Also open and read references/template-a4-pdf.md — you will need its chapter structure and page balance rules for Phase 2b and Phase 3.

2a. Interpret the Cards

Core Reading Principles

Combination-first: Never list individual card meanings in isolation. Read adjacent cards as overlapping pairs (1+2, 2+3, 3+4...) that modify each other. The meaning of a card shifts based on its neighbors.
Noun + Modifier model: The first card is the subject (noun), the second card modifies it (adjective/verb/adverb). Rider + Clover = "lucky news arriving." Clover + Rider = "a fleeting visitor."
Use modifier types systematically: Each card has a modifier_type in references/lenormand-cards.md. When interpreting A+B, apply B's modifier type to A's core meaning. Example: A=Heart (love), B=Mountain (blocking modifier) → "Love blocked or delayed."
Check for negative modifiers: 8 cards (Clouds, Snake, Coffin, Scythe, Whip, Mountain, Mice, Cross) darken adjacent cards. When one of these is the second card in a pair, the tone shifts darker regardless of the first card's polarity.
Neutral tone: Stay faithful to Lenormand's traditional meanings. Do not push toward positivity or negativity — the tone emerges from the cards themselves. Avoid empty reassurance and unfounded pessimism.
Honesty: The value of a reading lies in honesty, not in making the querent feel good or afraid.
Center dominance: In odd-numbered linear spreads, the center card is the core theme.
Full-length interpretation: The saved interpretation should be written at PDF-ready length — each chapter aiming for 4-5 paragraphs + takeaway (for 3-card/5-card spreads) or equivalent depth. The interpretation JSON stores the full version. Social card generation will compress this saved content.

Card Knowledge

Load references/lenormand-cards.md for each card's complete meaning (core, extended, career, love, health, person in both EN/ZH), keywords, polarity, modifier type, timing, and the systematic combination methodology. The modifier type system tells the agent HOW each card changes its neighbor — use this to synthesize pair meanings, rather than improvising from isolated keywords. Card data is sourced from lenormand-oracle (MIT License).

Spread-Specific Interpretation Frameworks

1-Card Yes/No:

State the verdict clearly (YES/NO/MAYBE)
Interpret the card's meaning applied to the question
Note any subtle nuances the binary verdict might miss

2-Card:

Read the two cards as a single combination (1+2)
Give the overall meaning as one connected sentence

3-Card / Daily:

Chain adjacent pairs: 1+2, then 2+3
Synthesize all three into one connected message
End with a practical takeaway or "今日提醒" (daily reminder)

5-Card:

Identify the center card (position 3) as the core theme
Chain pairs: 1+2, 2+3, 3+4, 4+5
Read all 5 as one connected narrative trajectory
End with practical guidance

9-Card Box Spread (3×3): Interpret from three perspectives:

Time Axis (columns) — Left column (1,4,7) = Past, Center column (2,5,8) = Present, Right column (3,6,9) = Future
Three Layers (rows) — Top row (1,2,3) = Conscious, Middle row (4,5,6) = Reality, Bottom row (7,8,9) = Subconscious
Cross Method — Core (position 5), Cross arms (2,4,6,8) = direct influences, Corners (1,3,7,9) = background/hidden factors
Key Combinations — Highlight notable card pairs and clusters
Overall Synthesis — All 9 cards woven into a unified reading

A-or-B Choice:

Chain-read Option A's 3 cards (1+2, 2+3)
Chain-read Option B's 3 cards (1+2, 2+3)
Compare the two trajectories side by side
Offer a tendency or caution, but respect that the final choice belongs to the querent

Output Format

Structure the interpretation as clear chapter sections with titles in the detected language. Use the chapter structure defined in references/template-a4-pdf.md for your spread type.

2b. Save the Reading

Phase 2b is mandatory and must complete before Phase 3 begins. Persist the reading in two parallel files under output/readings/:

1. Structured JSON — {reading_id}.json

Follow the schema defined in references/reading-schema.md. The JSON must contain:

version: "3.0"
reading_id: "{date}-{topic_slug}" (e.g., 2026-06-20-写作方向)
meta: language, spread_type, spread_label, question, date, generated_at, source, topic_slug
cards: array of card objects with position, card_id, names, polarity, position_label
interpretation: summary + chapters array, each chapter with id, title, short_title, and sections (paragraph / combination / takeaway)
verdict: (yesno-1 only) verdict + verdict_zh
optionA / optionB: (choice only) option texts + card arrays

2. Human-readable Markdown — {reading_id}.md

A readable copy including:

markdown

# {question}
**{spread_label}** · {date}

{summary}

## 抽出牌面
| 位置 | 牌 |
|------|-----|
| {position_label} | {number}. {name_zh} |

{for each chapter:}
## {chapter_title}

{paragraphs}

> **{combo_label}**: {combo_content}

> 💡 {takeaway}

3. Update the index — append to output/readings/index.json:

json

{
  "reading_id": "2026-06-20-写作方向",
  "language": "zh",
  "spread_type": "three",
  "question": "我的写作方向应该是什么？",
  "date": "2026-06-20",
  "generated_at": "2026-06-20T21:35:00+08:00",
  "topic_slug": "写作方向",
  "card_names": ["騎士", "幸運草", "書"],
  "products_generated": []
}

Confirm before continuing: Both {reading_id}.json and {reading_id}.md must exist on disk and the index must be updated. Only then proceed to Phase 3.

AGENT — PHASE 2 CHECKPOINT: Before proceeding to Phase 3, confirm the following. Emit this block filled with actual values.

text

PHASE 2 COMPLETE
─────────────────────────────
reading_id   : {reading_id}
json saved   : output/readings/{reading_id}.json — {EXISTS|FAILED}
md saved     : output/readings/{reading_id}.md — {EXISTS|FAILED}
index updated: output/readings/index.json — {EXISTS|FAILED}
chapters     : {N} chapters in interpretation
summary      : {first 80 characters of summary}...
─────────────────────────────
Proceeding to Phase 3.

If any field shows FAILED, stop and fix the save operation before continuing.

Phase 3: Generate Products

Hard gate — do not begin Phase 3 until Phase 2b is confirmed complete. Verify that output/readings/{reading_id}.json exists and output/readings/index.json has been updated. If either file is missing, complete Phase 2b first. Never generate products from in-memory interpretation data that has not been persisted.

Products are generated from the saved reading JSON. The user may request one or more product types. If no product type is specified, default to A4 PDF.

Seed-Based Workflow

Instead of building HTML from scratch, copy the seed HTML file from assets/, then fill in placeholders:

Copy seed to output/ with the appropriate filename
Replace {{PLACEHOLDER}} markers with data from the saved reading JSON:
- {{TITLE}}, {{DATE}}, {{SPREAD_LABEL}} etc. → from meta
- {{CARD_N_SVG}} → read SVG from cards/card-{NN}-{slug}.svg, strip <defs>…</defs>, paste inline
- {{CHAPTER_TITLE}} / {{CHAPTER_CONTENT}} → from interpretation.chapters[]
- {{SUMMARY}} → from interpretation.summary
Delete unused spread templates from the seed (keep only the one matching spread_type)
For chapters: duplicate the  block for each chapter
Render with the appropriate script

Product	Seed file	Render script	Output name
Product	Seed file	Render script	Output name
---------	-----------	---------------	-------------
`a4-pdf`	`assets/seed-a4-pdf.html`	`scripts/generate-pdf.js`	`lenormand-{spread}-{topic}-{date}.pdf`
`web-page`	`assets/seed-web-page.html`	— (self-contained)	`lenormand-web-{topic}-{date}.html`
`social-card`	`assets/seed-social-card.html`	`scripts/render-social-cards.js`	`lenormand-xhs-{topic}-{date}-NN.png`
`long-image`	`assets/seed-web-page.html`	`scripts/render-long-image.js`	`lenormand-web-{topic}-{date}.png`

3a. A4 PDF (Default)

Step 1 — Copy the seed file

bash

cp assets/seed-a4-pdf.html output/lenormand-{spread_type}-{topic_slug}-{date}.html

Do NOT write a new HTML file from scratch. You must copy the seed.

Step 2 — Fill all placeholders Open output/lenormand-{spread_type}-{topic_slug}-{date}.html. Replace every {{PLACEHOLDER}} marker using str_replace or sed, sourcing all values from the saved reading JSON at output/readings/{reading_id}.json. Required replacements:

{{TITLE}} → meta.question
{{DATE}} → meta.date
{{SPREAD_LABEL}} → meta.spread_label
{{SUMMARY}} → interpretation.summary
{{CARD_N_SVG}} for each card → inline SVG from cards/card-{NN}-{slug}.svg (strip outer <defs>…</defs>, inline inner content only, no nested <svg><svg>)
{{CHAPTER_TITLE}} / {{CHAPTER_CONTENT}} per chapter → from interpretation.chapters[]

Step 3 — Delete unused spread templates Remove all spread template blocks from the HTML except the one matching meta.spread_type.

Step 4 — Duplicate chapter blocks Find the  marker. For each chapter after the first, duplicate the chapter block. Chapters 2 and beyond must have style="break-before: page;" on their container element.

Step 5 — Render to PDF

bash

node scripts/generate-pdf.js output/lenormand-{spread_type}-{topic_slug}-{date}.html

Step 6 — Validate

bash

node scripts/validate.js output/lenormand-{spread_type}-{topic_slug}-{date}.html --type a4-pdf

Exit code must be 0. If non-zero, fix all reported failures and re-run Step 5 before continuing.

Step 7 — Record in index

bash

node scripts/manage-readings.js record-product {reading_id} --type a4-pdf

3b. Web Page

Step 1 — Copy the seed file

bash

cp assets/seed-web-page.html output/lenormand-web-{topic_slug}-{date}.html

Step 2 — Fill all placeholders Same process as 3a Step 2. Source all values from output/readings/{reading_id}.json.

Step 3 — Validate

bash

node scripts/validate.js output/lenormand-web-{topic_slug}-{date}.html --type web-page

Exit code must be 0. Fix any failures before continuing.

Step 4 — Record in index

bash

node scripts/manage-readings.js record-product {reading_id} --type web-page

3c. Social Card (1080×1440)

Step 1 — Load and compress interpretation Open output/readings/{reading_id}.json. For each chapter, produce a compressed version:

Max ~3 content blocks per card page (e.g. 2 combos + takeaway, or 3 short paragraphs + takeaway)
Shorten sentences, merge redundant points, cut filler
Never truncate mid-sentence — drop the weakest points instead
Estimated total block height per page must be ≤ 1100px at 30px font

Step 2 — Copy the seed file

bash

cp assets/seed-social-card.html output/lenormand-xhs-{topic_slug}-{date}.html

Step 3 — Fill all placeholders Same process as 3a Step 2, using compressed interpretation for chapter content.

Step 4 — Render to PNGs

bash

node scripts/render-social-cards.js output/lenormand-xhs-{topic_slug}-{date}.html

Step 5 — Validate

bash

node scripts/validate.js output/lenormand-xhs-{topic_slug}-{date}.html --type social-card

Exit code must be 0. Fix any failures (overflow, thin pages, card sizing) and re-run Step 4 before continuing.

Step 6 — Record in index

bash

node scripts/manage-readings.js record-product {reading_id} --type social-card

3d. Long Image (Full-Page PNG)

Generates a full-page PNG screenshot of the web page version. Useful for sharing on social media, messaging apps, or as a long single-image reading.

Workflow: Generate web page first (Phase 3b), then screenshot it.

Step 1 — Generate web page (ensure Phase 3b is complete) If the web page HTML does not already exist, run Phase 3b Steps 1–2.

Step 2 — Render to PNG

bash

NODE_PATH=/Users/1mether/.workbuddy/binaries/node/workspace/node_modules \
  /Users/1mether/.workbuddy/binaries/node/versions/22.22.2/bin/node \
  {{SKILL_ROOT}}/scripts/render-long-image.js output/lenormand-web-{topic_slug}-{date}.html

The script automatically derives the output path ({same_name}.png).

Step 3 — Verify output Check that the PNG file exists and has non-zero size:

bash

ls -lh output/lenormand-web-{topic_slug}-{date}.png

Step 4 — Record in index

bash

node scripts/manage-readings.js record-product {reading_id} --type long-image

Note: The full-page PNG inherits the web page's viewport width (720px). The height is determined by the page content length. For best results, run on a reading with all chapters filled. CSS @media print and page-specific styles are NOT applied — the PNG uses the screen layout.

Generating Multiple Products

User may request multiple product types at once: "生成 PDF 和网页版". Generate each as described above.

Product Regeneration (Post Hoc)

When the user references a past reading for a new product:

Identify the reading. Common triggers:
- "上次的解读" / "latest reading" → node scripts/manage-readings.js latest
- "列出最近的解读" → node scripts/manage-readings.js list --limit 5
- "用 #3 解读" / "第三张解读" → list then load by index
- "用 6月20号的解读" → search by date
Load the reading JSON from output/readings/{reading_id}.json
Ask which product type (if not specified in the user's request). Default to A4 PDF if the user only says "生成产品" without a type.
Run Phase 3 with the loaded reading and requested product type.

Bundled Resources

Resource	Purpose	When to Load
`references/lenormand-cards.md`	36-card database with MIT-licensed meanings (EN+ZH), modifier types, timing, negative modifier list, and systematic combination methodology	Phase 2 (always needed)
`references/spread-parsing.md`	Prompt parsing rules for all 6 spread types	Phase 1
`references/design-system.md`	Shared design tokens, fonts, colors, card slugs	Phase 3 reference
`references/template-a4-pdf.md`	A4 PDF chapter structure & page balance rules	Phase 3a reference
`references/reading-schema.md`	JSON schema for saved readings	Phase 2b (saving)
`assets/seed-a4-pdf.html`	Copy-and-fill A4 PDF HTML skeleton	Phase 3a (copy to output/)
`assets/seed-web-page.html`	Copy-and-fill web page HTML skeleton	Phase 3b (copy to output/)
`assets/seed-social-card.html`	Copy-and-fill social card HTML skeleton	Phase 3c (copy to output/)
`scripts/generate-pdf.js`	Puppeteer: HTML → A4 PDF	Phase 3a
`scripts/manage-readings.js`	CLI: list, find, update reading index	Phase 2b, post-hoc regen
`scripts/render-social-cards.js`	Puppeteer: social card HTML → 1080×1440 PNGs	Phase 3c
`scripts/render-long-image.js`	Puppeteer: web page HTML → full-page PNG	Phase 3d
`cards/*.svg`	Geometric Silence 36-card deck	Phase 3a, 3c (SVG card faces)
`cards/previews/*.png`	PNG previews of all 36 cards	Quick reference

Output Directory Structure

text

output/
├── readings/
│   ├── index.json                     # Reading manifest
│   ├── 2026-06-20-写作方向.json       # Structured reading data
│   └── 2026-06-20-写作方向.md         # Human-readable copy
├── lenormand-three-写作方向-2026-06-20.pdf    # A4 PDF product
├── lenormand-web-写作方向-2026-06-20.html     # Web page product
├── lenormand-xhs-写作方向-2026-06-20-01.png   # Social card product (page 1)
└── lenormand-web-写作方向-2026-06-20.png      # Long image (full-page PNG)

Deliverable Validation

Validation is run automatically via node scripts/validate.js at the end of each product generation step. See Phase 3a–3c for usage. All checks must pass (exit code 0) before presenting any product to the user.

Edge Cases & Notes

Missing universal prefix: Treat entire text as spread body if no --- delimiter
Daily reading: Auto-detected by 每日運勢; question is null, generate "daily fortune" context
9-card significator: Mark in card data if position 5 is noted as Significator
Card ambiguity: Resolve both Chinese (騎士) and English (Rider) case-insensitively
Freeform (non-Stargazer) input: If the input doesn't follow Stargazer format (no --- delimiter, no Stargazer ZH/EN markers), still run Phase 1 language detection — the CJK fallback in spread-parsing.md will determine the output language. All Phase 2–3 rules apply as normal.
Language consistency: Detect language FIRST from the prompt itself (via markers or CJK fallback). Do NOT override with user profile language preference. If Chinese prompt → all text in Chinese. English → all English. The interpretation must match the prompt's language, not the user's chat language.
Duplicate generation: If a product type already recorded in index, still regenerate (overwrite) but log a note
Social card chapter count: If a reading has only 1 chapter (yesno-1, two), produce a cover-only card (single 1080×1440 image)
Web page immutability: Web pages are standalone and self-contained — they don't link back to the reading data
Long image prerequisite: long-image depends on the web page already being generated (Phase 3b). If the user requests long image without specifying web page, silently generate both (web page first, then the screenshot).
Puppeteer Chromium sandbox: The renderer scripts keep Chromium's sandbox on by default (shared logic: scripts/chromium-args.js). --no-sandbox / --disable-setuid-sandbox are added only when running as root or when PUPPETEER_NO_SANDBOX=1 is set. If page.setContent() / page.goto() fail with "Requesting main frame too early!" in an environment where the sandbox cannot run, rerun with PUPPETEER_NO_SANDBOX=1 — do not hardcode the flags.
HTML escaping: When filling {{PLACEHOLDER}} markers in any seed template, HTML-escape user-derived text (the question, names, free-text notes) and generated interpretation prose — &, <, >, " — before substitution. The only unescaped substitutions are the template's own structural fragments (card SVGs, chapter markup the template defines).