ai-ppt-remix

---
name: ai-ppt-remix
description: Use when remaking an existing PPT or slide deck from a source PPTX and matching script or notes into a new deck of consistent AI-generated slide images, especially when one new slide should semantically merge one or more source slides while preserving important screenshots, text, and case-study evidence.
---

# AI PPT Remix

Use this skill when the user wants a source deck reworked into a new AI-visual deck without losing the original argument. This skill is for semantic slide remixes, not for generic "make me a new PPT" requests.

Also use the existing `PowerPoint` skill for rendering and rebuilding `.pptx` files, and the existing `imagegen` skill for per-slide generation.

## Transcript-first rule

When a real talk transcript, narration script, or spoken notes exist, they are the authority for semantic fidelity.

- The transcript is not background context; it is a required source.
- Do not simplify away transcript-specific claims just because the original slide only hinted at them.
- Numbers, dates, platforms, market comparisons, causal claims, and caveats mentioned in the transcript must be captured slide by slide as explicit constraints before image generation.
- A visually strong slide is still a failure if the transcript-specific point disappears into generic "marketing" visuals.
- If a slide's factual density is too high for a single clean AI image, stop and either switch to a hybrid rebuild flow or ask the user how much factual compression is acceptable.

## Evidence-fusion rule

When the user asks for AI-remixed slide images, source screenshots are evidence, not delivery tiles.

- Do not paste old slides, reference boards, or screenshot grids into the final image.
- Do not remove the evidence either. Rebuild it as large, recognizable visual evidence inside a new composition: charts become redrawn evidence charts, screenshots become stylized but identifiable artifacts, case studies keep their names, and spoken numbers become readable labels.
- Use "AI-fused evidence" when the user wants a new AI image but still needs key screenshots and facts to be visible. The final slide should feel newly generated while letting the speaker point to the same proof.
- A generated slide is not acceptable if the must-keep facts only appear as tiny background texture, unlabeled icons, or vague themed imagery.
- If the model cannot keep both visual quality and factual evidence in one pass, regenerate from a stricter manifest before changing the deck.

## When To Use

- The user has an existing `pptx` plus a matching script, narration, or detailed notes.
- One new slide should combine one or more old slides into a single semantic beat.
- Important screenshots, charts, labels, examples, or case-study evidence must survive the redesign.
- The result should be a visually consistent AI-generated deck, not a screenshot collage.

Do not use this skill for net-new decks with no source deck, or for simple edits that only need normal PowerPoint changes.

## Workflow

1. Prepare source materials.
- Render the source deck to preview PNGs with the `PowerPoint` skill.
- If there is already a target deck with a good visual direction, render that too and use it as the style-reference deck.
- Mark unchanged slides early. Reuse them instead of regenerating everything.

2. Group source slides by script meaning.
- One output slide equals one spoken idea, not one source slide.
- Group one or more source slides when they support the same script beat.
- Before generating, extract a fact checklist from the transcript for each group:
  - exact visible text that must stay readable
  - must-keep evidence screenshots or case-study visuals
  - must-keep transcript facts such as numbers, dates, platforms, comparisons, and causal links
  - any caveats or nuance that would be lost if the slide became generic
- Write a group manifest before generating. See `references/manifest-schema.md`.
- Treat this manifest as a contract. Do not generate first and backfill facts afterward.

3. Build one reference board per generated slide.
- Use `scripts/build_reference_boards.py`.
- Each board should show:
  - one style reference slide
  - all source slides in the semantic group
  - the relevant script excerpt
  - the visible text that must remain readable
  - the must-keep evidence
  - the must-keep transcript facts
  - any semantic remapping rules
- The reference board is only a prompt and review artifact. It must not be copied, tiled, or shrunk into the final slide.

4. Write prompts from meaning, not from layout.
- Use `references/prompting.md`.
- Ask for a brand-new 16:9 slide image.
- Require exact visible text where needed.
- List the must-keep screenshots, examples, labels, numbers, or diagrams explicitly.
- List transcript-derived facts explicitly, especially when they are easy for the model to blur away:
  - budgets, counts, dates, platforms, geography, ratios, and comparisons
  - who or what a number refers to
  - whether a point is an example, a benchmark, a warning, or a caveat
- Convert high-risk spoken facts into visible slide labels or callouts. If a fact matters to the talk, it should not depend on speaker memory alone.
- Crucial rule: place evidence according to what the script means, not according to the original pixel position.

5. Generate consistently across the deck.
- Generate one anchor slide first to lock palette, texture, density, and composition language.
- Reuse the accepted anchor slide as a style reference for later generations when possible.
- For high-risk pages, show both the reference board and the most relevant original slide.

6. Adopt approved images and rebuild the deck.
- Use `scripts/adopt_generated_slide.py` to crop and copy the chosen generated image into the working slide-image path.
- Rebuild the final deck with the `PowerPoint` skill.
- Preserve speaker notes and any unchanged slides.

7. Review semantically, not just visually.
- Reject any slide that becomes generic while dropping the source argument.
- Reject placeholder frames, fake browser chrome, or empty mockup boxes.
- Reject old-slide sticker layouts: a new background plus pasted source slide thumbnails is not an AI remix.
- Reject any slide that swaps a specific case study for a generic substitute.
- Reject any slide that loses transcript-specific facts or softens them into vague summary language.
- Reject any slide that keeps the title but drops the spoken point, such as the specific benchmark, spend comparison, or risk chain described in the transcript.
- Use a yes/no acceptance checklist per slide before adoption:
  - Is the spoken claim still visible?
  - Are the must-keep facts still present?
  - Are the must-keep screenshots or case studies still recognizable?
  - Would the speaker still be able to say the original lines naturally while this slide is on screen?
- For partial-deck repairs, hash or otherwise compare unchanged slide images before delivery to prove non-target slides were not modified.

## Non-Negotiables

- Do not shrink old slides and paste them into a new slide as tiny thumbnails for delivery.
- Do not let the model replace specific screenshots, examples, charts, or text evidence with generic filler.
- Do not preserve the original physical layout if the new slide structure changes the meaning.
- Do preserve the original semantic role of each example.
- Do not let transcript facts disappear just because they were spoken instead of typed on the source slide.
- Do not accept a beautiful slide that fails as speaker support for the actual talk.

Example:
- If the old slide used a different axis direction than the new slide, remap the examples to the correct new quadrants by label meaning, not by old screen position.

## Resources

- `references/manifest-schema.md`: group manifest fields and example JSON.
- `references/prompting.md`: prompt template, consistency rules, and semantic-fidelity checks.
- `scripts/build_reference_boards.py`: builds style-plus-source reference boards with script and must-keep notes.
- `scripts/adopt_generated_slide.py`: copies the selected generated image into the working slide image path with cover-crop resizing.