ppt-image

Installation
SKILL.md

PPT Image

This is the canonical PPT-generation skill. It replaces HTML/Beamer/tcbposter-first slide workflows with an image-native workflow: Codex plans the story, writes a strict visual prompt for every slide, generates each 16:9 slide image through gpt-image-2-skill, reviews the raster output, then packages the accepted images into a PPTX.

The backend is the installed gpt-image-2-skill CLI, normally at $CODEX_HOME/skills/gpt-image-2-skill/scripts/gpt_image_2_skill.cjs. Resolve the current skill directory from the loaded SKILL.md; commands below use $PPT_IMAGE_SKILL_DIR for this skill and $GPT_IMAGE_2_SKILL_DIR for the image backend. Do not use ppt-design, Beamer, tcbposter, Gemini, Paperbanana, or python-pptx layout reconstruction as the primary rendering route.

If shell commands are needed, set CODEX_HOME="${CODEX_HOME:-$HOME/.codex}", PPT_IMAGE_SKILL_DIR="$CODEX_HOME/skills/ppt-image", and GPT_IMAGE_2_SKILL_DIR="$CODEX_HOME/skills/gpt-image-2-skill" unless the skill is installed under different names.

Hard Boundary

Image generation is not a truth engine. Never ask GPT Image 2 to invent data, citations, bar heights, equations, or experimental numbers. Extract claims and numbers from source files first, lock them in a content manifest, and make the slide prompt reproduce only that manifest. For quantitative plots, use an existing plot as a reference image or create a plain data figure outside the image model, then ask GPT Image 2 to compose or restyle around it while preserving geometry. For formulas, keep only one short core formula per slide; if exact formula rendering fails after retries, stop and report the failure rather than silently accepting a wrong formula.

The default policy is reference-first. If the input contains real figures, plots, diagrams, screenshots, or paper assets, preserve them unless there is a clear reason to redraw. The image model may create surrounding slide layout, explanatory panels, and conceptual visuals, but exact author-provided figures should be inserted as original pixels whenever the slide's correctness depends on them. Use GPT Image 2 to generate a replacement figure only when no suitable source figure exists or when the user explicitly asks for a reinterpretation.

Output Contract

For every deck, create a self-contained project-local folder:

Installs
9
GitHub Stars
7
First Seen
Jun 12, 2026