GordenImage2PPTX — Slide Image to Editable PPTX
Reverse-engineers slide screenshots or any image page into a fully editable .pptx: text as real text boxes, frames and icons as movable image layers.
The skill mandates task-isolated RUN_ROOT directories, forbids scanning/reusing historical files, and enforces an imagegen-evidence manifest; data flow and external effects are largely transparent, scripts only read/write local files with no network exfiltration or credential handling. Deductions: no rollback/cleanup guidance, reliance on Codex built-in imagegen without security boundary notes, and unknown license/publisher attribution.
SKILL.md and references are layered and internally consistent; bundled scripts match documented parameters (chroma_key.py, compose_pptx.py), failure feedback is explicit (layout_guard blocks, edge_touch blocks delivery), and known failure modes plus remedies are documented. Deductions: static review without execution; no committed tests or CI covering key paths; actual success of imagegen generation and coordinate conversion is unproven; edge cases (icon slicing, drift) rely on manual retries.
Trigger conditions are clear in the frontmatter (image-to-editable-PPTX, icon extraction, etc.), inputs/outputs are well bounded, Chinese support is strong, and multiple runtimes are addressed. Deductions: image generation depends entirely on imagegen-class backends, limiting use in environments lacking them; boundaries for unsuitable inputs (hand-drawn/photo slides) are not declared.
Docs are well layered (overview, detailed references, schema) with examples and QA checklists and stable naming. Deductions: license metadata unknown, no versioning/changelog, unclear ownership and update path; the schema section in references/image-to-pptx.md is truncated mid-sentence, a documentation completeness defect.
By design the skill converts slide images into PPTX with editable text and movable image layers, offering real marginal value over manual rebuild, with multiple QA feedback loops. Deductions: output usability cannot be verified statically; the workflow demands multiple rounds of manual visual recalibration (icon, frame-anchor, text), so real cost is high and fidelity depends on imagegen quality, likely requiring notable rework.
The docs include concrete formulas, commands, and auditable QA artifacts, giving some traceability. Deductions: everything is author claims and prompt design; no third-party execution evidence, no CI workflows or committed test suites, and marketing-style superlatives are uncorroborated, so conclusions remain low-confidence inference.
- License and publisher metadata are unknown; verify authorization and attribution before commercial or enterprise adoption.
- Core image generation depends on Codex built-in imagegen-class backends; runtimes lacking them cannot complete key steps, and reachability of overseas image backends from mainland-China networks is unassessed.
- The workflow requires multiple manual calibration rounds (coordinate conversion, icon re-anchoring, frame-anchor re-anchoring); expect notable rework rather than one-shot high-fidelity restoration.
- This is a static-only review with no execution; no tests or CI corroborate key paths, so reliability conclusions are low-confidence.
- The schema section in references/image-to-pptx.md is truncated; treat SKILL.md as authoritative for layout contracts and report the defect to the maintainer.
What does this skill do, and when should you use it?
This is the reconstruction stage of the three-skill GordenSun/GordenSuperPPTSkills collection. It forces every input image — picture-format PPT pages, slide screenshots, any image page — into four layers: background, overall frame, icon/decoration elements, and text, then reassembles them into an editable .pptx at measured coordinates. Image layers are produced via extractive generation with an imagegen-class backend, while text is read by GPT's vision and written as genuine text boxes, with no traditional OCR. The README states the skill is Codex-only and has not been specifically adapted to other runtimes such as Claude.
Takes user-supplied slide images (one page at a time), creates an isolated RUN_ROOT directory per task; probes palette colors with probe_palette.py to pick a chroma-key backdrop; uses Codex's built-in image_gen with the current source image as the visible edit target to generate background, frame, and icon-sheet layers on a solid key color; strips backdrops with the bundled chroma_key.py and slices icon sheets with slice_grid.py; reads all normal text (position, size, color, weight, alignment) via GPT vision into layout.; runs layout_guard.py, placement_qa.py, and visual_compare_qa.py for coordinate-contract checks, source-box visual review, and side-by-side/blend/heatmap QA; finally compose_pptx.py assembles the .pptx with previews, and every page must emit an imagegen-assets-manifest. — pages lacking generation evidence fail and must be redone.
- Users who received slide screenshots or picture-format PPTs and want a truly editable PPTX to keep editing.
- Turning picture-format PPTs produced by GordenImagePPTGen into editable files for an end-to-end pipeline.
- Designers who need to extract icons, decorations, or stylized artwork out of a slide image.
- Users who want to replicate a slide's background and layout (including chart graphics) but only have an image, no source file.
- Users with multi-page image PPTs who need page-by-page reconstruction aligned to the original coordinates.
What are this skill's strengths and limitations?
- Output is not a screenshot — text is real editable text boxes; frame and icons are movable, replaceable, scalable image layers.
- Layered hard gates (generation-evidence manifest, layout_guard, placement_qa, visual_compare_qa) force redos instead of shipping defective pages.
- Each skill directory is self-contained with its own scripts/ and references/ — copy and use.
- Frame ships as one whole image by default, but can be split into individual parts on explicit request.
- README explicitly limits it to Codex — it depends on Codex's built-in image_gen and GPT vision and has not been adapted to other runtimes.
- Expensive: about 10% of a Plus 5-hour quota per image converted.
- No license file in the repo (copyright notice requires commercial use to credit the repo or author @Gorden Sun).
- No automated test suite; quality hinges on GPT vision and imagegen results, with no published success rates or benchmarks.
- Image layers are imagegen-generated; the frame is a single image by default rather than native shapes, so editing the frame means swapping the whole image.
How do you install this skill?
In Codex: send the repo URL https://github.com/GordenSun/GordenSuperPPTSkills to Codex and ask it to install the skill; or copy manually — cp -R GordenImage2PPTX "${CODEX_HOME:-$HOME/.codex}/skills/GordenImage2PPTX". Install dependencies: pip3 install python-pptx pillow numpy. The README does not document installation on other platforms (e.g., Claude Code).
How do you use this skill?
Codex only (README recommends GPT 5.5 with medium reasoning). Place the image in the current folder and prompt: "Use GordenImage2PPTX to convert XXX.png in the current folder back into an editable PPT, strictly following the skill steps." The skill walks each page through steps B0–B9; if you want the frame split into independent movable parts, say so explicitly in the prompt. The README warns on cost: converting one image consumes roughly 10% of a Plus subscription's 5-hour quota.
How does this skill compare with similar options?
Within the same repo, GordenSuperPPTSkill chains this skill with GordenImagePPTGen (one-shot "generate then convert"); GordenImagePPTGen only generates picture-format PPTs. For converting existing images alone, this skill suffices.