presentation — shared talk craft

A library. Engines load it. It does not scaffold a project.

You own the what (researched, grounded content) and the how-well (presentation craft). The engine owns the medium (Slidev markup or a Three.js world).

When to use

• slidedev-presentation or threejs-presentation is running.
• An engine needs Gate 1, the interview core, craft rules, or the storyboard gate.

When to skip

• The user asked to generate a deck. Load an engine instead.
• The user wants raw Slidev syntax. Use the official slidev skill.
• A one-off answer (a definition, a snippet). Just answer.

Non-negotiables (read before working)

These are what the user is paying for. Do not skip them because a deck "looks fine."

1. Research + ground BEFORE the interview. The moment you have a topic, run Gate 1 (references/research-grounding.md). Never build from model memory.
2. Build from the brief, not from memory. Every factual claim traces to the research/grounding brief.
3. Flag weak claims; don't assert them. If something is commonly repeated but poorly supported, skip it or caveat it. Never present a myth as fact.
4. Assertion-evidence + lean text. Titles are complete-sentence claims. The body is visual evidence, not a bullet dump. Keep on-surface text to a glance. Aim ≤ ~20 words of body per slide. Full sentences belong in speaker notes. (references/presentation-craft.md "Text budget")
5. One idea per slide; ~1–2 min/slide. Split rather than cram.
6. Fill the canvas. Every slide uses all available space with balanced proportions. Never top-anchor content above a dead band. Never float a small element in a big empty region. The engine has the layout recipe.
7. Code discipline. ≤5–7 lines visible at once, large, progressive reveal, dim irrelevant lines. Prefer a diagram over a wall of code. (references/technical-craft.md)
8. Generate every legitimately generatable asset. Placeholder only the un-generatable (real photos, product screenshots, demo videos) with a specific swap instruction. Never fabricate evidence. (references/assets.md)
9. Storyboard approval before writing slides (Gate 2).

What an engine must run

Step 0  Detect topic (and whether it is a specific library/tool)
Gate 1  Research & Ground   → references/research-grounding.md   (BEFORE interview)
Step 2  Interview core      → references/interview.md            (adaptive, one Q at a time)
Step 3  Deeper ground       → only if internals/architecture-focused or long-form
Gate 2  Storyboard approval → per-slide plan; user approves before any engine output

The engine inserts its own medium questions (Slidev theme, or Three.js world/theme/story/timer) around Step 2. Then the engine scaffolds, generates, verifies, and finishes.

Step 0 — Detect

Identify the topic. Decide whether it is also a specific library/tool (repo path / GitHub URL / package name / docs URL / names a concrete library / user is in a code repo). This turns on Gate 1 Track B in addition to Track A.

Gate 1 — Research & Ground (FIRST, before talking)

Load references/research-grounding.md. Run Track A (topic research, always) and, when applicable, Track B (code grounding). Light pass now (enough for a smart interview). Deeper pass later if the talk warrants. Produce the research + grounding brief.

Step 2 — Interview core

Load references/interview.md. Ask the canonical questions adaptively, one at a time, skipping anything the brief already answers.

Step 3 — Deeper grounding (conditional)

If the chosen areas are internals/architecture-focused or the talk is long-form, deepen the brief at the per-area depth the user set.

Gate 2 — Storyboard approval

Present a per-slide storyboard: assertion title · content/asset · layout or catalog form · speaker-note intent · place in the narrative arc · running time budget vs stated duration. Get approval/edits before writing any engine output (slides.md, src/talk.js, or world code).

Reference files