Phase 2: Rhetorical Architecture — Detail

Plan Analog Before Going Digital

Architecture decisions in this phase happen at the conceptual level — mode, opening pattern, narrative arc, sectioning, pattern strategy. These are best worked out away from slideware. Reynolds is emphatic: opening a presentation tool during planning prematurely commits the author to a template, a layout, and a default font when those decisions should still be fluid.

When the author is making architecture decisions in this phase, encourage analog tools — paper sketches, whiteboard diagrams, Post-it notes — for working through the structure before any slide is created. The five-step Reynolds workflow keeps slideware out of the picture for the first four steps: brainstorm → group/identify the core → analog storyboard → sketch visuals → only then transfer into slideware. Architecture decisions in this phase belong to steps 1–3; slideware belongs to Phase 5. See patterns/prepare/concurrent-creation.md for the full workflow.

The Joint Selection Process

This phase is a conversation, not a monologue. Use AskUserQuestion for each instrument selection. One decision per turn. Never combine multiple decisions into a single message — see the interaction-rules steering rule.

For each decision:

1. Extract the options from the vault summary (sections 2-13) and speaker profile (instrument_catalog). The vault is the living source — new instruments appear as more talks are parsed.
2. Present the options via AskUserQuestion with brief descriptions
3. Recommend based on the spec — put the recommended option first with "(Recommended)"
4. Wait for the author's choice before moving to the next decision

Mode Selection Logic

Read presentation_modes[] from the speaker profile. Each mode has a when_to_use field — use these to build a selection logic table dynamically. Present the modes with their descriptions and match signals from the spec.

Decision #2: Engine & Theme Sourcing

Pick the deck's tooling (pptx template vs presenterm terminal-markdown) and theme right after Mode — engine depends on the mode's typical_engine and constrains Slide Design, Template Patterns, and Illustration Strategy (presenterm has no pptx layouts; the illustration pipeline assumes pptx). Deciding it late forces re-litigation.

Source the choice via the shared wizard: idea-sourcing-wizard.md. Engine-specific bindings:

• Reads: profile → presentation_engines[] (roster + usage_count/out_of), the chosen mode's typical_engine, recent talks' outline.yaml engine, WebSearch (trends), author-supplied references.
• Presents: the multi-select source menu, proposals spanning the checked sources, then a single-select picking the final engine (+ theme pointer).
• Quick-default: resolve engine by precedence — chosen mode's typical_engine → highest-usage presentation_engines entry → pptx; set engine_source: "quick-default"; proceed immediately with a one-line note.
• Writes: talk.engine (closed enum pptx|presenterm), talk.deck_theme (free-string pointer), talk.engine_source (provenance).
• Summary-only / no-profile: present a flat two-option pptx/presenterm menu with hard-coded when-to-use; no usage tiers, no "your usual"/"series" sources; quick-default → pptx.

Theme stays a thin provenance pointer — the deck's look is owned by design_rules and, for illustrated decks, the illustration STYLE ANCHOR. Do not build a named deck-theme registry here.

Opening Pattern Selection Logic

Read instrument_catalog.opening_patterns[] from the speaker profile. Each pattern has a best_for field. Match to the spec's audience warmth, venue size, and context.

Narrative Arc Templates

Read instrument_catalog.narrative_structures[] from the speaker profile. Each has acts and time_allocation. Present the options with their time splits and best-for context.

Persuasive vs. Informative Architecture — Sparkline or Narrative Arc?

Before presenting narrative-arc templates, ask one upstream question: is this talk primarily persuasive or primarily informative?

• Persuasive = the audience is being asked to do or believe something different after the talk (sales pitches, strategic-direction announcements, fundraising, organizational change, advocacy keynotes, investor pitches).
• Informative = the audience needs to understand something but is not being asked to act on it differently (tutorials, technical deep-dives, scientific explanations, status updates, postmortems).

For persuasive talks, present sparkline (per patterns/build/sparkline.md) as the default top-level structural option. The sparkline's two named turning points (Call to Adventure, Call to Action) and "new bliss" close are purpose-built for moving audiences to action. Stack with one of the contrast-driven sub-structures inside the middle (problem-solution, compare-contrast, cause-effect, advantage-disadvantage).

For informative talks, present the existing narrative-arc templates (three-act and variants) as the default. The three-act structure suits content that needs to be understood; the sparkline is overkill and can feel manipulative when there's no genuine action to take.

The two patterns can coexist: an informative talk can have a small sparkline-shaped closing argument, and a persuasive talk can have informative sections inside its middle. But the choice of top-level structure matters because it shapes time allocation across the three sections — sparkline allocates ≤10% to "what is" baseline and most of the time to the persuasive middle; narrative-arc typically allocates ~25-50% to the middle, with longer setup and resolution.

When the mastery_and_novelty domain is available and classifies narrative-arc or sparkline as regular/signature, surface that authorized preference but do not let it override the persuasive-vs-informative diagnostic. Never infer the preference from raw tags or occurrence rows. A speaker accustomed to narrative-arc tutorials switching to a sales pitch should switch to sparkline for that talk; the architecture should match the talk's purpose, not the speaker's habit.

Action Typology — Pre-Plan the Call to Action

When sparkline is selected (or whenever the talk includes a call-to-action moment), pre-plan the audience's action diversity at the architecture level. Per patterns/build/call-to-action.md, every audience contains four action-temperament types — Doer (instigates activities), Supplier (provides resources), Influencer (changes perceptions), Innovator (generates ideas) — and the call-to-action must address at least one ask per type.

This is an architecture-phase concern (not a content-phase one) because the asks shape the entire backward-design of the talk: if you can't name a credible Doer ask, the talk lacks an actionable thesis; if you can't name an Influencer ask, you haven't accounted for audience members who can't directly execute but can spread the idea. Write the four asks before writing any other content; the rest of the talk is in service of making them feel earned.

Decision #11: Pattern Strategy

Read patterns/_index.md for the current taxonomy. Use only the catalog-history source selected by Phase 0's emitted history_source. When its status has history_enabled: true, authorize each derived lane independently through available_classification_domains / domain_available(domain). Never infer a domain from profile recency, schema tolerance, Section 15 prose, classification_fields_available: true, or opportunity_rows_available: true. Profile schema v4 and Section 15 v2 are occurrence-only; profile schema v5 and Section 15 v3 are policy-bound.

When mastery_and_novelty is available, present patterns in 4 tiers:

PATTERN STRATEGY for "{talk title}"
===================================
YOUR TOOLKIT (signature):
  ✓ Narrative Arc (22/24 talks) — recommended for this format
  ✓ Bookends (18/24) — strong with this audience
  ✓ Expansion Joints (20/24) — essential for 45→20 min adaptation

WORTH CONSIDERING (contextual):
  ○ Talklet (3/24) — good fit for the 20-min constraint
  ○ Foreshadowing (7/24) — pairs well with your arc style

NEW TO YOU:
  ★ [NEW] Preroll — display bio/topic on screen before you start
  ★ [NEW] Seeding the First Question — plant an easy Q for Q&A

SHAKE IT UP:
  ⚡ [WILD CARD] Red, Yellow, Green — audience voting with colored cards
  ⚡ [WILD CARD] Cave Painting — one giant canvas instead of slides

WARNINGS:
  ⚠ [CONTEXTUAL] Dual-Headed Monster — co-presented talk, define handoff points
===================================

Tier logic:

1. Signature — current-cohort mastery_levels.signature entries, always shown.
2. Contextual history — current-cohort mastery_levels.regular or mastery_levels.occasional entries matching the talk context.
3. New to You — current-cohort never_used_patterns / mastery_levels.never_tried, filtered by relevance and marked [NEW]. This set is exact: never use first detection in the newest talk or not_yet_observed as evidence of novelty.
4. Shake It Up — exactly 1–2 current-cohort never-tried options used as provocations, not prescriptions.

Current-cohort strengths requires mastery_and_novelty; underused_patterns requires underuse; signature_combinations requires signature_combinations; and by_mode requires modes. Pattern or antipattern movements, score/breadth trend, and score drivers require trends. Never infer one domain from another.

Antipattern warnings. Derive contextual warnings from the spec (co-presented → Dual-Headed Monster, dense content → Bullet-Riddled Corpse, new format → Shortchanged, etc.). Catalog recurrence filtering and [RECURRING] labels belong to Phase 4's emitted recurring_antipatterns output. Do not recreate them during architecture. Current-outline detections always receive [CONTEXTUAL] and remain available without history. Explicit source_lane: "non_pattern" guardrails remain independent of this catalog lane.

Mastery-domain-disabled / summary-only mode: Pattern taxonomy still works — patterns come from the reference files alone. Present a flat relevant-patterns list without the four history tiers, usage statistics, novelty claims, or strengths. Do not call taxonomy entries "new to you" merely because mastery history is unavailable. Other policy-bound domains may still authorize their own narrow claims; contextual antipattern warnings always apply. Section 15 v3 can restore only the domains listed by its complete explicit provenance contract. Section 15 v2 is occurrence-only narrative/audit context, not tier authorization.

Across profiles, a catalog/scoring generation change resets occurrence and classification comparison. Within one generation, a changed policy_semantic_sha256 resets classification comparison. Neither boundary is speaker improvement or regression.

Enhance decisions 3-10 with pattern cross-references as shared vocabulary: when recommending an opening pattern, reference the taxonomy ID; when selecting a narrative structure, note which Presentation Patterns it maps to (e.g., "problem-solution" = Narrative Arc + Triad).

Decision #12: Illustration Strategy (when applicable)

Not every talk needs generated illustrations — demo-heavy, data-heavy, or screenshot-driven talks may not. When the author wants AI-generated illustrations, delegate to the illustrations skill for the full collaboration (optimization priorities, format vocabulary, a priority-driven model shortlist, style proposals grounded in vault visual_style_history, a style × model exploration render, visual continuity devices):

Skill(skill: "illustrations")

The skill writes the approved style_anchor block into outline.yaml, then returns control to Phase 2. Continue with the next decision (or the architecture gate) once the skill returns.

Slide Budget Calculation

Read guardrail_sources.slide_budgets[] from the speaker profile. Match the spec's duration to the closest budget entry. Read pacing for WPM and slides/min targets.