presentation-skill

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: presentation-skill
description: "Build, edit, redesign, render, and verify polished editable PowerPoint `.pptx` decks from a prompt, structured `outline.json`, local data, or a saved workspace. Use for presentation and slide-deck generation, lab/clinical/scientific reports, board and investor decks, editorial briefs, charts/tables/figures, template-inspired redesign, geometry/readability QA, rendered visual review, and reproducible deck workspaces. Aliases: PowerPoint skill, PPTX skill, presentation generator, slide-deck generator, deck builder, powerpoint-deck-builder, pptx-skill."
---

# Presentation Skill

Create editable PowerPoint decks from source files. The model owns narrative,
evidence, and design judgment; repository scripts own deterministic rendering,
staging, and QA.

## Backbone

- Treat `outline.json`, planning files, local data, and figure scripts as the
  source of truth.
- Build with repository scripts. Do not write one-off inline `python-pptx` or
  `pptxgenjs` deck code.
- Fix source and rebuild. Do not patch generated `.pptx` files when source is
  available.
- Keep charts, tables, text, and layout objects editable where practical.
- Run QA and inspect rendered slides before delivering a deck.
- Never install dependencies during a deck task. Report a missing dependency.
- Scaffold a new topic from its own evidence and story. Do not clone another
  deck's slide sequence as a house style.

## Start Here

Read only the context needed for the current phase:

1. `DESIGN.md` for the compact design contract.
2. `references/outline_schema.md` for fields and supported variants.
3. `references/model_adaptive_workflow.md` for execution-profile selection.

Then load one task-specific reference:

- Saved/rebuildable deck: `references/deck_workspace_mode.md`
- Existing PPTX edit: `references/editing.md`
- Data/figure workflow: `references/reproducible_workflow.md`
- Style inspiration or screenshot/template matching:
  `references/style_reference_catalog.md`
- Structural diversity or topic-to-grammar routing:
  `references/composition_grammar_catalog.md`
- PptxGenJS renderer changes: `references/pptxgenjs.md`
- Fresh-eyes rendered QA: `references/visual_qa_prompt.md`

Do not preload the corpus, all preset descriptions, or every workflow
reference. Search or open the selected item on demand.

## Model-Adaptive Route

Use workload profiles as orchestration controls, not as different quality
definitions:

- `quality-first` (`sol`, `frontier`, `pro`): difficult/high-stakes decks,
  complex evidence, full rendered review, optional bounded scouts.
- `balanced` (`terra`, `standard`): default professional route, at most one
  useful scout, one focused render/repair loop.
- `fast` (`luna`, `draft`): short internal drafts, deterministic routing,
  render-free first pass, then one final render.
- `auto`: choose from the request; high-stakes or evidence-heavy work becomes
  quality-first, explicit rough drafts become fast, everything else balanced.

The profile changes context and delegation, not the editable source contract or
final QA bar. Future models should use the smallest prompt that passes real
deck evaluations; do not add model-specific process prose without evidence.

For a new saved workspace, emit a compact brief automatically:

```bash
python3 scripts/init_deck_workspace.py \
  --workspace decks/my-deck \
  --title "My Deck" \
  --style-preset executive-clinical \
  --user-prompt "Original request" \
  --agent-profile auto
```

Read `agent_brief.md` first. It contains the active profile, style route,
commands, and completion rubric. Keep `deck_start_packet.json` on disk for
audit/recovery; do not paste it into the active model prompt.

To regenerate only the brief:

```bash
python3 scripts/model_adaptive_workflow.py \
  --workspace decks/my-deck \
  --packet deck_start_packet.json \
  --agent-profile auto
```

## Choose The Workflow

### Quick Deck

Use for a one-off 5-10 slide deck when no future rebuild works