Image to Editable PPT
Rebuild slide screenshots, PDFs, and image-based PPT files into object-level editable PowerPoint decks — not a full-page screenshot pasted into slides.
SKILL.md discloses data flows (PaddleOCR, Codex OAuth, OpenAI-compatible API, limited to task-local page images/prompts/references), forbids sending credentials, and requires approval requests in restricted environments; however README recommends full-access mode, SKILL.md instructs not to frame required calls as unsafe, and consent is implied by the conversion request. Deducted for default external upload and approval-steering language.
CLI runtime is cleanly structured with a deterministic state machine, reset/diagnosis paths for failed pages, and concrete error messages; but static review cannot reproduce execution, and key page-reconstruction quality depends on LLM behavior not verifiable from source. Deducted for unverifiable core paths and multi-round LLM loop uncertainty.
Trigger description is precise (including Chinese phrasings), input/output and non-fit boundaries are clear; Baidu AI Studio OCR is mainland-China reachable. Deducted because core image generation depends on Codex OAuth/ChatGPT endpoints that may be unreachable from mainland networks without third-party relay configuration.
MIT license, multilingual README, CI-enforced changelog, well-layered references with single authoritative homes, candid known-issues disclosure; but unverified publisher, CLI version only 0.1.0, and maintenance/update path rests on a single author. Deducted for incomplete versioning and attribution evidence.
README shows before/after examples; structured manifest/finalize validation adds real value; but cost is very high (may exhaust a 5-hour quota, 2-3x cost of image PPT generation), 100% fidelity is explicitly not guaranteed, and static review cannot confirm outputs are directly usable. Deducted for limited evidence of direct usability.
Repository contains real CI workflows (compileall + unittest) and committed tests, but coverage centers on dispatch concurrency, formula rendering, and backend contracts, not the page-reconstruction critical path; showcase images are self-attested with no independent reproduction. Deducted for thin critical-path evidence.
- Consent for conversion is treated as consent for external OCR and image backend calls; declare local-only/confidential handling for sensitive material.
- README recommends full-access mode and approval gates are deliberately softened; run in a controlled environment and monitor each external call.
- Very high cost: a single page may take 10+ minutes and multi-page runs may exhaust quota; cost is 2-3x image PPT generation. Use lighter alternatives unless editability is essential.
- Image generation backend depends on Codex OAuth/ChatGPT endpoints which may be unreachable from mainland-China networks without a relay.
- 100% fidelity is not guaranteed and lower-tier models are unsupported; outputs need human review.
- Publisher identity is unverified by FollowSkills; this is a static review with no execution, so confidence is low.
What does this skill do, and when should you use it?
This is a Codex-oriented skill that rebuilds single images, multiple images, multi-page PDFs, or image-based PPTX files into object-level editable .pptx decks. Readable text is restored as native text boxes, simple geometry as PowerPoint shapes, and complex visuals as standalone image assets with provenance records. It drives a deterministic state machine (prepare → dispatch → record → finalize) through the editppt CLI, with multi-page inputs processed in parallel by page workers. Text sizing and positioning are driven by Baidu PaddleOCR-VL measurements rather than model eyeballing. It does not author new presentations from scratch, and conversions are slow and token-heavy.
Normalizes inputs into pages/page_NNN/source.png and runs editppt prepare to create a run directory with per-page text hints; single-page runs are rebuilt locally by the parent agent, multi-page runs are dispatched to page workers up to max_concurrent_pages (default 6). Each page produces manifest., page.pptx, preview.png, and validation.; editppt run record validates page outputs against the manifest, and editppt run finalize rebuilds the final .pptx in page order and runs deck-level validation (slide count, media relationships, asset hashes, notes hashes). Image generation/editing prefers Codex's built-in image_gen.imagegen, falling back via the editppt image CLI to Codex OAuth and then an OpenAI-compatible API; text correction calls Baidu PaddleOCR-VL when a token is configured, otherwise degrading to offline geometric detection. Speaker notes from .pptx inputs are copied verbatim to the output.
- Users who received slide screenshots or photos and want a deck where text and elements can be directly edited or moved
- Users with a scanned or image-based PDF courseware who need it restored as an editable multi-page PPTX
- Users holding an image-based PPTX who want to re-edit it while preserving original speaker notes
- Designers recreating a single-page visual (e.g., an infographic) while keeping text editable
- Codex users with multi-page batch conversion needs whose environment supports subagent dispatch
What are this skill's strengths and limitations?
- Output is genuinely object-level editable PPTX: native text boxes, simple shapes, and layered image assets — full-slide screenshot-plus-text-overlay is explicitly rejected as a fallback
- Text sizing and positioning driven by PaddleOCR measurements keeps same-level text sizes consistent, more stable than visual estimation
- The editppt state machine makes runs deterministic and resumable, with deterministic structural validation at record and finalize
- Supports multiple input types (images, PDF, image-based PPTX) and preserves PPTX speaker notes verbatim
- Clear image-backend fallback chain: built-in image_gen.imagegen → Codex OAuth → OpenAI-compatible API
- High token cost: a single page can take over 10 minutes, a 10-page deck may exhaust ChatGPT Pro's 5-hour quota, and conversion cost may be 2-3x generating an image-based deck
- Multi-page inputs strictly require an environment with subagent dispatch capability
- Recommended only for gpt-5.5 and above; effectiveness on weaker models is not guaranteed
- Does not guarantee 100% replication — some elements and text positions may be slightly offset
- Photos, illustrations, textures, and hand-drawn decorations can only be moved as whole image assets, not edited internally
- Under Codex approval modes the workflow is frequently interrupted; full access is recommended
How do you install this skill?
Tell your agent: "Install the image-to-editable-ppt skill from https://github.com/ningzimu/image-to-editable-ppt-skill". Alternatively install via CLI: npx -y skills@latest add ningzimu/image-to-editable-ppt-skill --skill image-to-editable-ppt --agent <agent-id> --global. The editppt CLI is installed automatically by the AI during execution (pipx install --force --editable <skill-root>/cli); verify with editppt doctor. The skill entry is at skills/image-to-editable-ppt/SKILL.md, MIT licensed.
How do you use this skill?
In Codex, invoke the skill with $image-to-editable-ppt, then paste or attach images, a PDF, or a PPTX to the chat, or provide a local path, e.g.: "$image-to-editable-ppt convert this image to an editable PPT" or "$image-to-editable-ppt convert deck.pdf to an editable PPT". The README recommends running Codex with full access permissions, since approval prompts will interrupt OCR, image generation, and subagent steps. On first run the AI will ask once for a free Baidu AI Studio PaddleOCR token (https://aistudio.baidu.com/account/accessToken); configuring it noticeably improves text fidelity, but the skill also runs without it at reduced quality. Output lands in output/image-to-editable-ppt/{job-id}/final/.
How does this skill compare with similar options?
Complementary to the same author's codex-ppt-skill: that skill generates brand-new decks from articles, reports, or ideas, while this one rebuilds existing visual slides into editable form. The author also notes that for tweaking just one or two slide images, directly using gpt-image's editing capability is a lighter-weight alternative.