---
name: presentation-intelligence
description: >
  Presentation Intelligence — the ChatGPT Agents Studio agent rebuilt on Claude skills.
  Research-grounded, audience-specific 16:9 native editable slides bound to the LLM Presentation Generation Standard.
  Use when making slides, pitch decks, weekly status, QBRs, board packs, or when the user mentions Presentation Intelligence, GPT agent, or Claude slide skills.
---

# Presentation Intelligence (Claude rebuild)

This is the ChatGPT agent **Presentation Intelligence** (`agt_6a944dcc8d2c8191b08890892489aee1`) rebuilt so the **renderer is a Claude skill**, not ChatGPT canvas / image-slides / python-pptx one-shots.

Binding spec: `docs/LLM-Presentation-Generation-Standard.md` (v1.0, 2026-08-30).

## What changed from the GPT studio agent

| GPT studio | This rebuild |
|---|---|
| Skills row empty | This skill + Claude `quality-rules` + `frontend-slides` density + native PPTX |
| Apps / web search | Optional; never invent sources |
| 1 file: LLM Presentation Generation Standard | Same file, in-repo, enforced in code |
| ChatGPT writes slides | Claude skill writes `SlideSpec`; slide-deck.io exports native `.pptx` |

Do **not** flatten slides to images. Do **not** one-shot prompt-to-deck. Do **not** finish on “Thank you.”

## Compact generator instruction (Standard §22)

Build the presentation to change a specific audience outcome, not to fill a template. First define the audience, delivery mode, decision or learning goal, central takeaway, credible evidence, objections, and constraints. Research every non-trivial claim and maintain a claim ledger. Give every slide exactly one communication job and one primary claim. Use takeaway titles; connect claim, evidence, implication, and transition. Select visuals for explanatory value. Produce native editable PowerPoint objects. Reject factual errors, overflow, generic imagery, repetitive layouts, placeholders, and flattened output.

## Pipeline (Standard §4) — run these stages

1. Brief interpreter
2. Clarification (only consequential gaps; else record assumptions)
3. Evidence / claim ledger (sourced lines only)
4. Audience model
5. Narrative (project update / decision / research / …)
6. Slide job planner — one job + one claim per slide
7. Visual director — Claude quality-rules + frontend-slides density
8. Copy editor — takeaway titles, notes that do not repeat the slide
9. Layout engine — mixed native layouts
10. Renderer — `runSlideAgent` / `buildWeeklyDeck` / `buildResearchDeck` → `SlideSpec`
11–15. Critic, fact check, accessibility, file integrity, PIS-100 (automated subset in `lib/agent/pis-score.ts`)

## Claude skills this renderer uses

| Skill | Job |
|---|---|
| This skill | Audience contract, jobs, claim ledger, PIS gates |
| `quality-rules` | 6–16 word takeaway titles, layout mix, bullet cap, dark punctuation, closing ask |
| `frontend-slides` | Density limits, no scrolling slide, split overflow |
| Anthropic `pptx` contract | Text/chart/table stay objects |

## How to generate

Prefer in this order:

1. ConnectedWork (Jira / Notion / Asana / Linear / ClickUp / Smartsheet CSV / Project XML)
2. Pasted notes / transcript / brief

Then:

```
POST https://slide-deck.io/api/agent/slides
```

or, in this repo, `runSlideAgent()` in `lib/agent/run.ts`.

## ChatGPT Agents Studio

On `chatgpt.com/agents/studio/edit/agt_6a944dcc8d2c8191b08890892489aee1`:

1. Skills → Add skill → this SKILL.md (`https://slide-deck.io/agent/presentation-intelligence/SKILL.md`)
2. Replace generator behavior with: call `createSlides` from `https://slide-deck.io/agent/openai-actions.json`
3. Keep the Standard file attached
4. Never let ChatGPT draw image-only slides when editability is promised

Paste-ready instructions: `https://slide-deck.io/agent/gpt-instructions.md`

## Anti-patterns (Standard §21)

Topic-label titles. One-shot decks. Card-grid sameness. Invented metrics. Flattened PPTX. Agenda-then-topics. Thank-you closer. Notes that repeat the slide. Pie charts for precise comparison. Pop-color psychology.