Sermon Prep Architecture

A Process Design for AI-Assisted Expository Preaching

Status: Draft for Josh's review — correct what's wrong, fill what's missing
Based on: Psalm 23 project context, Art of Teaching Workbook, Crouch/Kim formation framework, past conversation summary

1. Philosophy: What This System Is and Isn't

This is a formation-first sermon prep process. AI serves as a research companion and structural scaffolder — never as the voice, the pastor, or the source of spiritual insight.

Three governing tensions (drawn from Crouch/Kim):

• Presence over performance. The sermon is an act of pastoral care, not content production. AI should slow the work down, not speed it up past the point of formation.
• Mirror, not person. AI reflects back structure, patterns, and research. It cannot hear from God, feel the weight of the congregation, or carry the vulnerability of the preacher. It offers raw materials and scaffolding — never scripts.
• The allness test. Every AI interaction gets measured: does this develop heart, soul, mind, and strength for loving God and neighbor? If it shortcuts any of those, it gets cut.

What AI May Do

• Study aids, word studies, structural observations on the text
• Source synthesis and cross-referencing from provided materials
• Structural scaffolding (movements/chunks/seams drafts)
• Language polishing and options (never final voice)
• Cultural biopsy research (what competing stories are out there?)
• Observation prompts and formation questions
• Administrative support (newsletter, calendar, logistics)

What AI Must Not Do

• Simulate pastoral presence, spiritual direction, or "hearing from God"
• Write the sermon (scripts, manuscript drafts, "impact-optimized" content)
• Replace the slow, silent, embodied work of sitting with the text
• Substitute for human feedback (Frank, Eric, the preaching team, the congregation)
• Generate vulnerability, stories, or personal conviction on Josh's behalf

2. Reusable Assets (Carry Across Every Sermon Project)

These are posture and method documents — they don't change with the text.

Reusable Principles (not files, but baked into the process)

• Josh's voice first. Always cite the preacher's own words, memos, and observations before any external source or AI-generated content.
• Expository bias. Observe the text before you move it. Structure, verbs, imagery, literary context — then interpretation, then application.
• Formation over information. The goal is not what happens in the sermon but what happens through the sermon in a people (Comer, Session 9).
• Cultural biopsy, not culture war. Take a piece of culture, hold it under a microscope, calmly analyze it (Rieff method via Comer). Where does it come from? Where is it going? Is it at odds with Jesus?
• One clear takeaway. "If you remember one thing…" — every sermon gets one (from the Psalm 23 project's movements_chunks_seams.md).

3. Sermon-Specific Build (Rebuilt for Each Text)

These are created fresh for each sermon but follow a consistent structure.

4. The Process (Phases, Not Steps)

These phases overlap and loop. They're not strictly sequential — but the order of priority matters. Phase 1 always comes first. AI involvement increases as you move through the phases.

Phase 1: Marinate (AI: Off)

Who: Josh + God + silence
When: Weeks before preaching. The earlier, the better.
What happens:

• Sit with the anchor text. Read it slowly, repeatedly, in multiple translations.
• Pray. Wait. Don't fill the silence.
• Record stream-of-consciousness voice memos — raw, unedited, wherever you are. (Creek, car, room, walk.)
• Notice what the text stirs up in your own life. What are you afraid of? What are you hoping? Where does this text find you?
• Pay attention to what's happening in the life of the congregation. Who are you preaching to?

AI's role: None. This is solitude, silence, and the slow work of the Spirit. An AI companion here defeats the purpose entirely (Crouch: "it robs you of learning not to need to speak to God").

Output: Raw voice memos. A felt sense of where the sermon might live.

Phase 2: Observe (AI: Light Touch)

Who: Josh + text + commentaries + AI as study partner
When: 10–14 days before preaching.
What happens:

• Expository observation of the text: structure, movements, repeated words, surprising verbs, imagery, grammar (who acts? to whom? toward what end?).
• Literary context: what comes before and after in the letter? How does this pericope function in Paul's argument?
• Canonical echoes: where else does this language/imagery appear? (e.g., Exodus 34 → Colossians 1)
• Word studies on key terms.
• Read 1–2 commentaries on the passage.

AI's role:

• Surface structural observations Josh might have missed.
• Cross-reference canonical echoes.
• Summarize or highlight relevant sections from provided commentaries/sources.
• Ask observation questions: "Did you notice that the verb shifts here?" "This word only appears in X other places."
• Always defer to Josh's observations first. AI adds, it doesn't lead.

Output: Expository Map (draft). Updated Sources list.

Phase 3: Listen (AI: Moderate)

Who: Josh + sources + formation community + AI as research partner
When: 7–10 days before preaching.
What happens:

• Go deeper into sources: books, podcast transcripts, other sermons on the text.
• Talk to people — Frank, Eric, the team. What are they seeing in this text? What's the series trying to do?
• Revisit voice memos from Phase 1. What's still alive? What's faded?
• Begin identifying the theological center: what is this text doing to the reader/hearer?
• Start noticing cultural stories that compete with the text's claims.

AI's role:

• Synthesize provided source materials (highlight relevant sections, surface connections).
• Help with cultural biopsy: "Here are the dominant narratives about [topic] your congregation is likely carrying."
• Cross-reference the series arc: how does this week's text relate to what Eric preached last week?
• Surface formation questions: "What practice could train people in what this text describes?"
• Never override Josh's instincts about what the sermon needs. Offer raw materials, not direction.

Output: Enriched Expository Map. Cultural biopsy notes. A growing sense of the sermon's theological center and emotional register.

Phase 4: Build (AI: Active Partner)

Who: Josh + AI as structural scaffolder
When: 5–7 days before preaching.
What happens:

• Lay out the Movements: Hook, Head, Heart, Hands (or whatever variation fits).
• Fill Chunks: what story goes where? What quote? What observation? What illustration?
• Craft Seams: 1–3 sentence transitions between movements and chunks.
• Identify the "If you remember one thing…" takeaway.
• Run an appreciation pass: is there at least one moment of specific, sensory, relational warmth? (VAK check.)
• Decide on sermon notes format: outline vs. in-depth outline vs. manuscript.

AI's role:

• Draft structural scaffolds (movements/chunks/seams) based on Josh's voice memos, expository map, and instincts.
• Offer 2–4 options for seams, not one "best" version.
• Suggest chunk ordering possibilities.
• Polish language — tighten sentences, clarify transitions, suggest stronger verbs.
• Run the appreciation pass: flag where the draft feels cold or abstract.
• Keep options modular and editable. Never deliver a finished script.

Output: Movements/Chunks/Seams document (working draft). Sermon notes in Josh's preferred format.

Phase 5: Refine (AI: Selective)

Who: Josh + human feedback + AI for specific asks
When: 2–4 days before preaching.
What happens:

• Get human feedback — Frank, Eric, a trusted friend. Read it aloud. Use the feedback roster: tell each person what type of feedback you need from them specifically (Comer, Session 5). One person for exegetical accuracy, another for emotional clarity, another for structural flow.
• Edit for clarity, flow, and conviction. Cut what doesn't serve the one thing.
• Practice delivery. Notice where you stumble — those are the seams that need work.
• Pray over the sermon and the congregation.
• Final formation check: does this sermon form people toward love of God and neighbor, or does it just inform them?

AI's role:

• Specific, targeted asks only: "Help me tighten this transition." "Is there a better word for X?" "Does this illustration land?"
• Run the editorial checklist from pastoral_guardrails.md.
• AI does not replace human feedback. Frank's pushback > Claude's polish.

Output: Final sermon notes. Readiness to preach.

Phase 6: Recover, Then Harvest (AI: Off, Then Light Touch)

This phase has two distinct beats, in strict order. Comer (Session 7) is insistent: you don't jump to analysis before your body, soul, and spirit have been tended to. Preaching requires your whole self, and that expenditure needs intentional recovery — not an immediate performance review.

Beat 1: Recover (AI: Off)

Who: Josh + rest + God
When: Immediately after preaching through the next 1–2 days.
What happens:

• Body first. Relax. Eat good food. Sleep. Don't touch the sermon.
• Soul second. Do something regenerative and fun for your personality — something that has nothing to do with ministry.
• Spirit third. Tend to your relationship with God apart from the sermon. Pray, but not about preaching. Be with God as a person, not a preacher.

AI's role: None. This is embodied recovery. The Crouch/Kim framework applies here as much as it does to Phase 1 — the temptation to immediately analyze, optimize, and extract lessons is exactly the "effortless power" impulse that erodes formation.

Beat 2: Harvest (AI: Light Touch)

Who: Josh + feedback community + AI as archivist
When: Later in the week after preaching, once recovery is complete.
What happens:

• Craft feedback loop. Listen to or watch the sermon. Gather input from your post-sermon feedback roster — each person speaking to their area of strength. Note what worked, what fell flat, what surprised you.
• Record a post-sermon voice memo. Capture reflections while they're still warm but after recovery has settled your nervous system.
• Archive reusable material. Write harvest entries as markdown in the sermon's archive/ folder — craft lessons, exegetical connections, reusable illustrations, structural patterns, post-sermon reflections. These files are the source of truth.
• Index into the knowledge store. Run python -m rag index sermons/ so the new entries are searchable from future sermons. (See §6: Knowledge Store.)
• Connect forward. How does what you learned this week inform next week's sermon or the rest of the series arc?

AI's role:

• Help organize and archive reusable material from the sermon.
• Write harvest entries as clean, titled markdown so they index well.
• Run python -m rag index sermons/ after writing harvest files.
• Process feedback into concrete, actionable notes (only if invited — never unsolicited critique).
• Human feedback is the primary input. AI organizes what others have said; it doesn't substitute for their perspective.

Output: Archived sermon materials. Feedback notes. Harvest entries indexed into the knowledge store. A preacher who's rested, not just productive.

5. Formation Checkpoints (Built Into the Process)

At three points, the process pauses for a formation gut-check:

1. Before Phase 2 begins (Observe): "Have I spent enough time in Phase 1? Am I rushing to production because silence is uncomfortable?"
2. Before Phase 4 begins (Build): "Is the theological center something I believe and am being formed by — or is it just something I've assembled from sources?"
3. Before Phase 5 ends (Refine): "Does this sermon form people, or just inform them? Would I preach this if only ten people showed up?"

These are not AI-administered checkpoints. They're questions Josh asks himself — or that Frank or a trusted friend asks him.

6. Knowledge Store

Past harvest entries are stored as markdown files and indexed for semantic search. The markdown files are the source of truth. The vector index (.qdrant/) is gitignored and rebuildable.

Commands

python -m rag index sermons/                      # Index new/changed files (incremental — skips unchanged)
python -m rag search "theme or concept"            # Top 5 results (default)
python -m rag search "theme or concept" --top 10   # Override result count
python -m rag stats                                # Check what's indexed

When to use

• Phase 6 (Harvest): After writing harvest files to the sermon's archive/ folder, run python -m rag index sermons/ so they're searchable for future sermons.
• Phases 2–5: When the preacher asks for connections to past work, run python -m rag search "query" and surface results. Do not surface results unsolicited.

First-time setup

pip install -r requirements.txt

First index run downloads embedding models (~600MB, cached after that).

7. Project File Structure (Per Sermon)

teaching/
├── CLAUDE.md                            # System-level posture + knowledge store rules
├── sermon_prep_architecture.md          # This document
├── .venv/                               # Project venv (rag package installed here)
├── .qdrant/                             # Vector index (gitignored, rebuildable)
├── _reusable/                           # Carry across every sermon
│   ├── pastoral_guardrails.md
│   ├── spiritual_formation_and_ai.md
│   ├── appreciation_template.md
│   └── feedback_roster.md
├── _template/                           # Copy to start a new sermon project
│   └── ...
└── sermons/
    └── col-1-15-20-cosmic-christ/       # Example sermon project
        ├── README.md                    # Series context, preaching date, team notes
        ├── anchor_scripture.md          # Full text of the passage
        ├── expository_map.md            # Text movements, verbs, imagery, theology
        ├── sources.md                   # Curated bibliography for this text
        ├── movements_chunks_seams.md    # Sermon architecture (working draft)
        ├── cultural_biopsy.md           # Competing narratives and formation contrasts
        ├── voice_memos/                 # Raw stream-of-consciousness transcripts
        │   ├── memo_01.txt
        │   └── memo_02.txt
        └── archive/                     # Post-sermon: final notes, harvest entries
            ├── final_sermon_notes.md
            └── harvest-*.md             # Indexed into knowledge store

8. The Handoff Principle

Every AI interaction in this system ends with a handoff. Claude's last move is always some version of: "You own tone, direction, and delivery. Here's what I've laid out — take what serves you, leave the rest, and go be with your people."

The sermon doesn't belong to the system. It belongs to Josh, to the congregation, and to the Spirit who works through the foolishness of preaching.

This document is a draft. Josh: mark it up. Tell me what's real, what's aspirational, what's missing, and what's wrong. I'd rather be corrected now than build on a false foundation.