k2f-research-deck

K2F Research Deck — AI-Native Aesthetic Design

Purpose

Teach agents to blueprint research decks before layout: one scientific claim per slide, traceable evidence, explicit caveats before the close, and no numbers or figures redrawn from memory.

Built on K2F

This skill produces research presentations as .K2F packages. Follow the k2f skill for authoring: unpack or init a workspace, edit the semantic tree, pack/verify, and export. K2F keeps agents on meaning (slide roles, evidence, visual intent)—not coordinates—while the engine compiles layout, typography, and theme into consistent, projector-ready output.

Before you build, install the core skill: npx skills add suzheng/k2f --skill k2f.

When to Use

• Preparing a lab meeting or PI update from a paper, preprint, or experiment log.
• Building a grant or internal review deck that must prove novelty without overclaiming.
• Converting dense manuscripts into a 6–15 slide talk with faithful metrics and figures.
• Rebuilding generic “paper summary” AI slides into a skimmable evidence story.

Workflow

Step 1. Generate the content_and_design.md

Single blueprint for semantic content and aesthetic intent. Use labeled blocks per slide (role, headline, body, visual type, asset notes) that the k2f skill can map later.

Sub-steps

1. Brief frame — Audience, decision target (inform, approve next step, defend novelty), time budget, language. Live talk = sparse bullets; async deck = self-explanatory headlines per slide.

2. Source-of-truth audit — List abstract, methods sketch, figures/tables, exact results, citations, limitations. Tag NEEDS USER INPUT for gaps; never invent stats, n, or citations.

3. Evidence spine — Arc before titles: stakes → gap → question/hypothesis → design → results (strongest first) → interpretation → caveats → next step/ask. External reviewers: contributions early; lab updates: lead with freshest result.

4. Slide architecture (include / cut) — Target 6–15 slides; merge or omit when inputs are thin:

   • Title (paper metadata) → Context/problem → Gap/question → Approach overview → Methods (diagram-led) → Main results (split quant vs qual if both exist) → Mechanism/interpretation → Limitations → Next steps → (optional) key references.
   • Cut literature surveys, duplicated abstract+introduction, method prose that belongs in backup, and extra ablation slides that do not change the decision.
   • Split when two claims share one slide (problem+hypothesis, multiple unrelated figures, or results+interpretation).

5. Per-slide content contract — For each slide specify:

   • Headline: conclusion, not section label (“Knockdown reduces latency 42% at n=12” not “Results”).
   • Body: max 3–5 bullets or one visual block; ~2 lines per bullet; no orphan single-word lines.
   • Visual type: methods pipeline, results chart, comparison matrix, qualitative grid, equation focus, contribution list, agenda/outline—prefer chart/table/figure intent over paragraphs for data.
   • Asset fidelity: source figure/table (figure #, caption) vs schematic only; do not bulletize tables that should display as tables.

6. Numbers — One unit system; matching metrics across slides; n/CI/test on hero numbers; [VERIFY] if uncertain.

7. Design intent — Tone: restrained, evidence-led. Hierarchy: title largest; one hero figure/metric/equation per slide. Density: one message per slide; section breaks between movements. Palette: calm academic, one accent, projection-safe contrast.

Rules & constraints

• One-claim gate: if the headline needs “and,” split the slide or demote a thread to backup.
• Fidelity gate: quantitative slides use user-supplied numbers or explicit TBD—not rounded or “typical” values.
• Figure gate: main plots/tables use source assets or explicit chart intent—not prose-redrawn data.
• Caveat gate: limitations or confounds before the closing ask.
• Novelty gate: one scannable “what’s new”; cut known background.

Common failure modes

• Ten slides of background, one of results → reorder spine; promote results and design earlier.
• Methods as paragraph walls → one pipeline diagram + 3 bullets max.
• Garbled or invented statistics → revert to source table/chart intent or mark TBD.
• Missing limitations → add explicit caveats slide; avoids PI “what about confound X?” derail.
• Duplicate paper abstract on multiple slides → single context slide, then gap.
• Crowded multi-panel figures → one panel per slide or callouts in the blueprint.
• Overclaiming headlines → match data; speculation under “open questions.”

Quality checklist

• Decision target and audience stated at top of the blueprint.
• Every slide has one headline claim and a declared visual type.
• Metrics and assets trace to inputs, TBD, or [VERIFY].
• Caveats and next steps before the final slide.
• 6–15 slides (or justified); design intent covers tone and hierarchy for the deck.

Step 2. Use the content_and_design.md and k2f skill to generate a K2F file and export to an appropriate file format.

Follow k2f to instantiate slides from the blueprint, apply roles from design intent, verify, and export the presentation-ready deliverable.