Illustrations

This skill is an action router — pick the step that matches the user's intent and execute only that step. Do not run other steps; do not parallelize.

Step 1 inspects the request and the talk-directory state to decide the mode (Strategy / Generation / Thumbnail) and which subsequent step is the entry point. The "Multi-mode chaining" section at the end of Step 1 is the one explicit exception, and only triggers when a single invocation requests multiple modes.

Owns every AI-generated image the toolkit produces: deck illustrations, build chains, and thumbnails. Reads the vault for visual history, outline.yaml for the style anchor and slide-level prompts, and the speaker-profile.json for visual_style_history and publishing_process.thumbnail config.

The auto-loaded steering rules are the constitution: illustration-rules (edit vs regenerate, build chains, iteration hygiene), title-overlay-rules (safe-zone composition), and thumbnail-generation-rules (Phase 7 specifics). Do not restate them here — apply them.

Key Files & References

Step 1 — Route by Mode

Determine which of three modes applies and execute only the matching steps:

• Strategy — outline has no STYLE ANCHOR yet, or the author wants to revise it. Run Step 2 (freshness), then Steps 3-9 (style strategy), then stop unless generation was also requested.
• Generation — outline has a STYLE ANCHOR and per-slide prompts. Run Step 2 (freshness), Step 10 (illustrations), Step 11 (builds, if any slides have a builds: block), and Step 12 (apply to deck, if a .pptx exists).
• Thumbnail — talk has been delivered and a video URL is available. Skip to Step 13.

Multi-mode chaining

If — and only if — a single invocation requests multiple modes (e.g., "design the visual style, then generate everything"), run them in order Strategy → Generation → Thumbnail. Proceed immediately to the first applicable step; do not pause for confirmation between modes. A single-mode invocation runs exactly the one matching step's chain and stops.

"Proceed immediately" governs handoff between modes — it never overrides the within-Strategy human gates (Steps 3, 4, 9). Those stop for the speaker even in a chained invocation.

Step 2 — Check Image-Model Freshness

Run the deterministic precheck first, then report its verdict in one line — never skip this step silently:

python3 skills/illustrations/scripts/model_registry.py --check-freshness

The script emits JSON: last_reviewed, age_days, stale, and the full models roster (ids, aliases, attributes). State the verdict before proceeding (e.g. "Registry reviewed 2026-06-02, 0 days old — fresh").

• stale: true — use WebSearch to identify the current flagship image models from Google (Gemini image, Imagen) and OpenAI (gpt-image-*), plus any other vendor with a public image API. Reconcile the registry in skills/illustrations/scripts/model_registry.py: add new flagships, drop discontinued ones, refresh the attribute tiers, bump REGISTRY_LAST_REVIEWED. "Nano Banana" is Google's codename for the Gemini image line (Nano Banana Pro = Gemini 3 Pro Image) — fold a codename into the matching entry's aliases, never add it as a separate model. The speaker approves the edits.
• stale: false — the roster is current; do not web-search.

For Generation mode entering an existing outline, also confirm the outline's baked model is still in the roster. Check it against the JSON models list (the alias map resolves codenames). If absent or clearly superseded, surface that to the speaker — they keep the baked model or re-run the exploration (Step 3). Never silently swap the model.

generate-illustrations.py dispatches by family: gemini-* / nano-banana-* (Google generateContent), imagen-* (Google :predict), gpt-image-* (OpenAI). A model from a new vendor family needs a model_family() + _call_<vendor> adapter before it can render — surface that as a follow-up script change.

Proceed immediately to Step 3 or Step 10 per Step 1's routing.

Step 3 — Source Style Ideas

Open the strategy collaboration by asking where the visual ideas come from. Present an AskUserQuestion multi-select of idea sources — your usual, mode / series match, new to you, wild card, what's trending, "I'll drive (here are my references)" — plus a Quick-default fast path. The 3-4 proposals in Step 7 span the checked sources. Never skip this step silently; never bake a model or style from this menu by reasoning — every candidate reaches the speaker as a rendered sample in Steps 8-9.

Source vocabulary, the Quick-default path, no-profile degradation: skills/illustrations/references/strategy.md.

Proceed immediately to Step 4.

Step 4 — Establish Optimization Priorities

Elicit what the speaker optimizes for with an AskUserQuestion multi-select (checkboxes, not radio): cost, speed, quality, build-editability. Auto-add build-editability when any slide has a Builds: block. Never skip this step silently — the priorities drive the shortlist in Step 6.

Proceed immediately to Step 5.

Step 5 — Define Format Vocabulary + Composition

Ask the speaker — AskUserQuestion, never infer — how titles and footers are rendered:

• Bleed — title + footer baked into every image, stylized to the art (the example noir deck). Striking and consistent, but not editable after generation. Sets style_anchor.composition: poster-theatrical and locks every illustrated slide to FULL (no IMG+TXT, no safe zones — EXCEPTION/screenshot slides without an image_prompt are exempt); the text treatment + footer are recorded in the anchor at Step 9.
• Overlay — titles + footers added by PowerPoint over a per-slide safe zone. Editable, uniform font, less integrated. Standard composition; the format vocabulary is FULL / IMG+TXT / EXCEPTION + any talk-specific additions.

Detail: skills/illustrations/references/strategy.md.

Proceed immediately to Step 6.

Step 6 — Narrow the Model Shortlist

Filter the roster to a shortlist by the Step 4 priorities — no render yet:

python3 skills/illustrations/scripts/model_registry.py --shortlist <priorities>

The roster is a seed cache, not an allowlist. To rank a model not in it (a new flagship from Step 2, or one the speaker names), WebSearch its attributes and pass --add '<json>', or list its id directly in candidates.json — see strategy.md.

Proceed immediately to Step 7.

Step 7 — Propose Styles Across the Checked Sources

Propose 3-4 style options spanning the Step 3 sources, grounded in concept fit + vault context (the speaker's visual_style_history, rhetoric-style-summary.md Section 13). These are candidates for the render, not a commitment — the speaker picks from rendered pixels, not prose. Option template: skills/illustrations/references/strategy.md.

Proceed immediately to Step 8.

Step 8 — Render the Exploration Grid

Write style-explore/candidates.json (styles × shortlist × formats per the schema), then render:

python3 skills/illustrations/scripts/generate-illustrations.py \
  <outline> --style-explore style-explore/candidates.json

This writes the grid under style-explore/, an index.md contact sheet, and a rendered.json manifest of what actually rendered. The grid is the only place a model becomes eligible to bake. Never skip this step silently. The speaker picks a style + model from style-explore/index.md.

candidates.json input + rendered.json output contract: skills/illustrations/references/style-explore-candidates-schema.md.

Proceed immediately to Step 9.

Step 9 — Bake the Anchor, Then Verify

Write the style_anchor block — chosen model + per-format anchors + visual continuity devices (conventions) — into outline.yaml. For poster-theatrical composition, also set style_anchor.composition: poster-theatrical, style_anchor.embedded_footer: <text>, and style_anchor.text_treatment: <how baked title + footer are rendered> (e.g. "glowing hand-script neon on an in-scene surface").

Anchor vs per-slide split — everything that must look identical across slides lives in the anchor; only the scene and the literal title vary per slide:

• Anchor: visual style (per-format anchors), text_treatment (the title + footer rendering style), and embedded_footer (the full footer text). These keep every baked title/footer consistent.
• Per slide: image_prompt carries only the scene; text_overlay carries only the literal title string for that slide — never the text styling.

Then run the deterministic precheck and report its verdict in one line; never skip this step silently:

python3 skills/illustrations/scripts/generate-illustrations.py \
  <outline> --check-style-explore

It confirms the baked model was rendered in the grid (exit 0) or refuses with an actionable message (exit non-zero). On failure the baked model was never shown — pick a rendered model from style-explore/index.md and re-verify, or re-render (Step 8). Step 10 generation re-runs the same gate, so a baked-but-unrendered model cannot generate.

Full protocol — the option template, continuity options, the gate contract: skills/illustrations/references/strategy.md.

Proceed immediately to Step 10 if generation was also requested; otherwise finish here.

Step 10 — Generate Deck Illustrations

Batch-generate every missing slide illustration from the outline:

python3 skills/illustrations/scripts/generate-illustrations.py \
  outline.yaml remaining

Review with the author. For targeted corrections use --fix (preserves the near-good output); for additions use full regeneration; for removals use --edit. The edit-vs-regenerate asymmetry rule (illustration-rules §1) governs which to pick. Save iteration versions (v2, v3) instead of overwriting — see illustration-rules Iteration Hygiene.

Operational detail (compare modes, prompt patterns, retry ladder): skills/illustrations/references/generation.md.

Proceed immediately to Step 11.

Step 11 — Generate Builds

If any slides in the outline have a builds: block, generate the backwards-chained build images. Each step's input is the previous step's output — never regenerate independently from prompts. The edit prompt for each step comes from builds[].erase (the additive builds[].desc is the human-facing reveal); --build skips any non-final step whose erase prompt lacks a "Keep ..." clause.

python3 skills/illustrations/scripts/generate-illustrations.py \
  outline.yaml --build all

Output: illustrations/builds/slide-NN-build-MM.<ext> where <ext> is the MIME-derived extension for each step (.jpg / .png / .webp depending on the model and source image). Build-00 is the empty frame; build-N is the full image. Detail and the per-step contract: skills/illustrations/references/builds.md.

If no slides specify builds, proceed silently to Step 12.

Step 12 — Apply Illustrations to Deck

Insert generated illustrations and build sequences into the .pptx. Build slides replace their parent slide rather than duplicating after it; speaker notes go on the final build step only.

The script contract is DECK ILLUSTRATIONS_DIR OUTLINE_YAML (positional, in that order), with optional --out, --image-ext, --scrim-color, --scrim-alpha. It writes a new <stem>-with-titles.pptx next to the input deck unless --out is given.

python3 skills/illustrations/scripts/apply-illustrations-to-deck.py \
  deck.pptx illustrations/ outline.yaml

If no .pptx exists yet (Phase 5 hasn't run), finish here — presentation-creator Phase 5 will call back into this skill at Step 12 once the deck is built.

Step 13 — Generate Thumbnail

Run the thumbnail composition for a delivered talk. Surface 3–5 candidate slides ranked by visual impact, let the speaker pick, then compose:

python3 skills/illustrations/scripts/generate-thumbnail.py \
  --slide-image illustrations/slide-NN.png \
  --speaker-photo "$SPEAKER_PHOTO" \
  --title "HOOK TITLE" \
  --aesthetic <photo|comic_book>

Aesthetic precedence (thumbnail-generation-rules §7): explicit speaker preference → default_illustration_style → confirmed intents → photo. For illustrated decks, also pass --portrait-style "<anchor>" so the portrait is pre-stylized to match the deck.

Iteration is conversational — change one thing at a time (style variant, expression, colors, title text, slide). Detail: skills/illustrations/references/thumbnails.md.

Finish here.