compose

דבר תורה — Orchestrator (v4.2)

Architecture

This skill is an orchestrator — it dispatches specialized agents in parallel, then feeds results to the writer. It does NOT do research or writing itself.

Phase 1 — Research (parallel, haiku)
  ├── previous-analyzer        ← scan output/ for patterns
  ├── source-researcher        ← Sefaria lookups
  ├── historical-researcher    ← if topic.type == "research" or orientation == "modern"
  └── source-researcher        ← mussar sources (if mussar requested)

Phase 2 — Analysis + Verification (parallel)
  ├── philosophical-analyzer   ← sonnet: deep reasoning with Phase 1 sources
  └── source-verifier          ← haiku: verify all Phase 1 citations

Phase 3 — Writing (opus)
  └── torah-writer             ← opus: compose final piece from all results

Config Input

config:
  language: "he" | "en" | "bilingual"
  orientation: "philosophy" | "kabbalah" | "hasidut" | "hazal" | "modern" | "custom"
  format: "dvar-torah" | "post" | "shiur"
  length: "short" | "medium" | "long" | "shiur"
  topic:
    type: "parsha" | "philosophical" | "research"
    value: "..."
    sub_topic: "..." | null
  thinkers: [...]
  context:
    type: "birth" | "brit_milah" | "bar_bat_mitzvah" | "wedding" | "shabbat" | "holiday" | "shiva" | "yahrzeit" | "general"
    description: "from contexts-guide.md"
  previous_patterns:
    mode: "avoid" | "base_on"
    base_documents: [...]
    template_guidance: {...}
  include_mussar: true | false

Orientation Routing

For full thinker profiles, works, and Sefaria references, see references/thinkers-guide.md. The table below is a quick routing summary only.

Orientation	Key Thinkers	Primary Sefaria Texts
philosophy	רמב״ם, רס״ג, רלב״ג, קרשקש, אלבו, ר״י הלוי, אבן עזרא	Guide for the Perplexed, Emunot veDeot, Milhamot HaShem, Or Hashem, Kuzari
kabbalah	זוהר, רמ״ק, האר״י, רמח״ל	Zohar, Pardes Rimonim, Etz Chaim, Daat Tevunot
hasidut	בעש״ט, מגיד, תניא, ברסלב, קדושת לוי	Tanya, Likkutei Moharan, Keter Shem Tov
hazal	תנאים, אמוראים, מדרשים	Talmud Bavli, Jerusalem Talmud, Bereishit Rabbah, Midrash Tanchuma
modern	קאופמן, קסוטו, פינס, אלטמן + ANE texts	(historical-researcher agent)

Sefaria category hints

• kabbalah: get_text_or_category_shape(path="Kabbalah")
• hasidut: get_text_or_category_shape(path="Chasidut")

Execution

Step 0: Interactive Wizard (if needed)

If the user did not provide full config — invoke wizard skill to collect parameters. Once config is complete, proceed to orchestration.

Step 1: Resolve topic

If topic.type == "parsha" and no specific parsha given:

mcp__claude_ai_Sefaria__get_current_calendar()
# or faster: python scripts/sefaria_api.py get-calendar

Set topic.value from the calendar result.

---

<use_parallel_tool_calls> If you intend to call multiple tools and there are no dependencies between the tool calls, make all of the independent tool calls in parallel. Prioritize calling tools simultaneously whenever the actions can be done in parallel rather than sequentially. Maximize use of parallel tool calls where possible to increase speed and efficiency. However, if some tool calls depend on previous calls to inform dependent values, do NOT call these tools in parallel.

Agent dispatch pattern:

• Phase 1 agents: launch ALL in a single message with background: true
• Phase 2 agents: launch BOTH in a single message with background: true
• Phase 3 (torah-writer): foreground — wait for completion </use_parallel_tool_calls>

PHASE 1 — Research (parallel, all background)

Launch ALL of these agents in a single message with background: true:

Agent 1: previous-analyzer (haiku)

Agent(
  subagent_type: "dvar-torah:previous-analyzer",
  model: "haiku",
  background: true,
  prompt: """
    MODE: {previous_patterns.mode}
    {if mode == "base_on": "SELECTED_FILES: {previous_patterns.base_documents}"}
    Scan output/divrei-torah/ and return analysis per your instructions.
  """
)

Agent 2: source-researcher (haiku)

Agent(
  subagent_type: "dvar-torah:source-researcher",
  model: "haiku",
  background: true,
  prompt: """
    TOPIC: {topic.value} ({topic.type})
    ORIENTATION: {orientation}
    THINKERS: {thinkers}
    CONTEXT: {context.type}

    Search Sefaria for sources relevant to this topic.

    IMPORTANT: When the source is Talmudic or has Steinsaltz commentary available,
    FIRST read Steinsaltz for pshat understanding:
      get_text("Steinsaltz on <reference>")
    This gives you the baseline meaning before philosophical analysis.

    For orientation={orientation}, focus on:
    - philosophy: Guide for the Perplexed, Emunot veDeot, Milhamot HaShem, Or Hashem, Kuzari
    - kabbalah: Zohar, Pardes Rimonim, Etz Chaim, Daat Tevunot
    - hasidut: Tanya, Likkutei Moharan, Keter Shem Tov
    - hazal: Talmud Bavli, Bereishit Rabbah, Midrash Tanchuma
    - modern: academic scholarship + ANE parallels

    Search at least 3 different books. Find connections between sources.
    Return structured YAML with sources, quotes, and sefaria_refs.
  """
)

Agent 3: historical-researcher (haiku, CONDITIONAL)

Only launch if topic.type == "research" OR orientation == "modern":

Agent(
  subagent_type: "dvar-torah:historical-researcher",
  model: "haiku",
  background: true,
  prompt: """
    TOPIC: {topic.value}
    VERSE/TEXT: {topic.sub_topic or topic.value}

    Research the historical and cultural context of this text.
    Find ANE parallels, archaeological evidence, and classical commentary.
    Return structured YAML.
  """
)

Agent 4: mussar source-researcher (haiku, CONDITIONAL)

Only launch if include_mussar == true:

Agent(
  subagent_type: "dvar-torah:source-researcher",
  model: "haiku",
  background: true,
  prompt: """
    TOPIC: {topic.value}
    ORIENTATION: mussar
    TASK: Search ONLY mussar literature for sources related to this topic.

    Search in:
    - Mesillat Yesharim
    - Chovot HaLevavot
    - Shaarei Teshuvah
    - Shemonah Perakim

    Identify which middah/virtue is relevant and find 2-3 mussar sources.
    Return structured YAML with sources, quotes, and practical applications.
  """
)

Wait for all Phase 1 agents to complete before proceeding.

---

PHASE 2 — Analysis + Verification (parallel)

Launch BOTH agents in a single message with background: true:

Agent 5: philosophical-analyzer (sonnet)

Agent(
  subagent_type: "dvar-torah:philosophical-analyzer",
  model: "sonnet",
  background: true,
  prompt: """
    TOPIC: {topic.value}
    ORIENTATION: {orientation}
    THINKERS: {thinkers}
    CONTEXT: {context.type}

    SOURCES FROM PHASE 1:
    {source_researcher_results}

    {if historical: "HISTORICAL CONTEXT: {historical_researcher_results}"}
    {if mussar: "MUSSAR SOURCES: {mussar_researcher_results}"}

    Build a philosophical analysis:
    1. Identify the core philosophical question
    2. Map each thinker's position using the sources provided
    3. Build critical dialogue between positions
    4. Synthesize conclusions

    If sources are insufficient, you may search Sefaria for additional texts.
    Return structured YAML with question, positions, dialogue, synthesis.
  """
)

Agent 6: source-verifier (haiku)

Agent(
  subagent_type: "dvar-torah:source-verifier",
  model: "haiku",
  background: true,
  prompt: """
    CITATIONS TO VERIFY:
    {all citations from Phase 1 agents}

    Verify each citation through Sefaria MCP.
    Format all references uniformly.
    Build bibliography: תנ״ך → חז״ל → ראשונים → אחרונים → מחקר
  """
)

Wait for both Phase 2 agents to complete.

---

PHASE 3 — Writing (opus)

Launch the writer with ALL collected results:

Agent 7: torah-writer (opus)

Agent(
  subagent_type: "dvar-torah:torah-writer",
  model: "opus",
  prompt: """
    WRITE a {format} in {language}.

    CONFIG:
    - Format: {format}
    - Length: {length}
    - Language: {language}
    - Orientation: {orientation}
    - Context: {context.type} — {context.description}
    - Thinkers: {thinkers}

    RESEARCH BUNDLE:
    ================

    PREVIOUS ANALYSIS:
    {previous_analyzer_results}

    SOURCES:
    {source_researcher_results}

    {if historical: "HISTORICAL CONTEXT:\n{historical_researcher_results}"}

    {if mussar: "MUSSAR:\n{mussar_researcher_results}"}

    PHILOSOPHICAL ANALYSIS:
    {philosophical_analyzer_results}

    VERIFIED CITATIONS:
    {source_verifier_results}

    {if base_on: "TEMPLATE GUIDANCE (FOLLOW STRICTLY):\n{template_guidance}"}

    INSTRUCTIONS:
    1. Read references/templates.md for the {format}/{length} template
    2. Read references/contexts-guide.md for {context.type} guidance
    3. Use ONLY verified citations (status: verified or corrected)
    4. Write the complete piece
    5. Save to: output/divrei-torah/{format_dir}/YYYY-MM-DD-{topic_slug}.md
    6. Include frontmatter with all metadata
  """
)

This agent runs in FOREGROUND — wait for it to complete.

---

Context Guidance

Read references/contexts-guide.md per context.type:

• shiva: comforting tone, themes of loss and meaning
• wedding: joyful and poetic, themes of partnership and Shekhinah
• bar_bat_mitzvah: encouraging, themes of responsibility and free choice
• general: follow orientation only

base_on Mode

If previous_patterns.mode == "base_on" and template_guidance exists:

• Pass template_guidance.strict_instructions to torah-writer
• Writer MUST follow: same structure, register, citation density, length (±20%)

Output

Final output saved by torah-writer to:

output/divrei-torah/
  ├── dvar-torah/YYYY-MM-DD-<topic>.md
  ├── posts/YYYY-MM-DD-<topic>.md
  └── shiurim/YYYY-MM-DD-<topic>.md

Frontmatter:

---
date: YYYY-MM-DD
format: dvar-torah | post | shiur
length: short | medium | long | shiur
language: he | en | bilingual
orientation: philosophy | kabbalah | hasidut | hazal | modern | custom
context: general | wedding | shabbat | shiva | ...
parsha: שם הפרשה (if relevant)
topic: הנושא
thinkers: [רמב״ם, רס״ג]
sources_count: N
agents_used: [source-researcher, philosophical-analyzer, torah-writer, ...]
---

Sefaria MCP Tool Priority (for agents)

Research agents use tools in this order:

1. english_semantic_search — primary discovery
2. get_links_between_texts — connections & commentary
3. get_text — exact text (or: python scripts/sefaria_api.py get-text "<ref>")
4. get_text_catalogue_info — metadata
5. get_text_or_category_shape — structure
6. get_topic_details — topic context
7. search_in_book — focused book search
8. search_in_dictionaries — term definitions
9. text_search — general fallback

This orchestrator uses only get_current_calendar (Step 1) or python scripts/sefaria_api.py get-calendar.