deepPPT
A research-style PowerPoint generation and QA skill that turns papers and outlines into consistent academic decks, with scripted layout checks and rendered visual review before delivery.
Evidence shows strong least-privilege design: Baidu OCR is off by default, requires task-scoped grant/revocation, auto-revokes on completion; ImageGen availability is decoupled from user intent; credentials are never persisted; output cannot overwrite sources; fingerprint binding prevents reusing stale reports. Deducted for unverified publisher identity, the fact that OCR still uploads user documents to a third party despite the consent flow, and confirmation relying on CLI discipline where no approval UI exists.
Real three-platform CI plus extensive committed unit tests covering lifecycle, permissions, idempotency and fingerprint checks constitute auditable execution evidence, exceeding the static cap of 10. Deducted because nothing was executed in this review, test files are partially truncated, failure-feedback quality on abnormal input is only partially inferable, and visual review depends on host vision or humans, so determinism cannot be fully verified.
Target scenarios (lab meetings, journal clubs, defenses, bilingual academic decks) and trigger conditions are explicit; platform routing boundaries (no native-fidelity claims on WSL/Linux) and non-fit ranges are disclosed; Chinese examples are complete; Baidu OCR is reachable from mainland China. Deducted for heavy multi-dependency local installation (PowerPoint/LibreOffice/Node), limited server/CI usability, and no usage evidence for trigger precision.
Well-layered docs (SKILL.md, bilingual READMEs, style guide, patterns, schemas, examples), version 0.2.0, MIT license, pyproject packaging with CI asserting SKILL.md ships in the wheel. Deducted for no standalone changelog, anonymous 'contributors' authorship, unverifiable publisher identity, and update path dependent on a personal repository.
Atomic delivery of an editable PPTX only after dual gates (deterministic QA + rendered visual review) is a clear value claim, and CI actually rebuilds and validates example PPTX files, surpassing prompt-only alternatives. Deducted because real-task output quality is unverified in a static review, visual review depends on host/human vision, and heavy local dependencies hurt cost/benefit for lightweight users.
Multiple auditable evidence types exist: CI workflow, committed test suite, downloadable example PPTX files, fingerprint-bound reports. Deducted because independent multi-source reproduction is absent, test evidence is only partially visible (truncated files), and no third-party verification exists, capping at 5.
- This assessment is a static source review only; nothing was executed and confidence is low.
- Full functionality requires local PowerPoint, LibreOffice, Node.js and other dependencies; servers and headless environments only support generation and preflight.
- Baidu OCR uploads user document images to a third-party network service; users with sensitive material should enable it cautiously despite the grant/revoke mechanism.
- Publisher identity is unverified; maintenance and updates depend on a personal repository, creating long-term uncertainty.
- Visual review quality depends on host vision or human inspection; a text-only model cannot complete that gate.
What does this skill do, and when should you use it?
deepPPT is an Agent Skill that generates Chinese or English research-style PowerPoint decks for lab meetings, defenses, journal clubs, and conferences, delivering a new editable .pptx. It converts error-prone steps into scripted checks: environment verification, PPTX text extraction, package integrity checks, layout QA, and local PowerPoint/LibreOffice render previews. Every deck follows one fixed visual grammar — a two-block title bar, navy-to-light segmented rule, navy chevron labels, and a red key-finding box reserved for decisive results only. Run state persists in a run directory and can be resumed after interruption. Input PPTX files are treated as evidence and never modified in place.
Reads PDFs, outlines, bullet points, or existing PPTX files and extracts their content; records a deck plan conforming to a JSON schema where every slide has an action title, bounded claim, evidence, and source references; builds a new PPTX with PptxGenJS 4.0.1; runs deterministic QA (qa_layout.py blocking duplicate template lines, red boxes on agenda pages, red-box overuse, visible provenance labels, etc.) and renders 150 DPI slide previews; requires a visual review (host vision or explicit human) to pass before atomically copying the reviewed revision to the output path. Optionally calls Baidu OCR for image-only sources via configure/ocr commands, gated by explicit per-task user permission with no tokens stored.
- A graduate student turning a paper PDF into a 12-slide Chinese lab meeting deck with rendered QA and a final PPTX.
- A PhD candidate preparing defense slides that need the unified template look and a red key-finding box only on the conclusion page.
- Weekly group reporting: quickly rebuilding an English journal club deck in fixed style from an outline or bullet points.
- Anyone who wants an existing PPTX's content extracted and rebuilt as a new deck without touching the original file.
- A conference speaker who needs Chinese or English academic slides checked page-by-page for overlap, clipping, and contrast before delivery.
What are this skill's strengths and limitations?
- Fixed visual grammar keeps decks generated by different agents looking like one template family.
- Dual delivery gates: a run completes only after deterministic QA and rendered visual review both pass, with reports bound to the exact PPTX hash.
- Resumable runs — state persists under the run directory and can be continued after interruption.
- Network image upload (Baidu OCR) requires explicit per-task opt-in; only a permission boolean is stored, never credentials.
- MIT licensed, with downloadable Chinese and English example PPTX files and previews.
- Native PowerPoint rendering only works in a signed-in interactive desktop session on macOS/Windows; WSL/Linux is limited to LibreOffice fallback.
- Intro-page concept images are only available via Codex host ImageGen and require an explicit user request; all other runtimes skip that path.
- The full toolchain is heavy: Python 3.12, Node.js, pptxgenjs, LibreOffice, and on Windows also PowerPoint desktop.
- No public automated test suite; version numbers disagree between README (v0.1.3) and SKILL.md (v0.2.0).
How do you install this skill?
The skill folder must be named deep-ppt. macOS/Linux: git clone https://github.com/jiadizhunine/deepPPT.git ~/.codex/skills/deep-ppt (use ~/.claude/skills/deep-ppt for Claude Code); cd into it, then python3.12 -m venv .venv && .venv/bin/python -m pip install -r requirements.txt; then npm install -g [email protected]. On macOS, optionally: brew install node [email protected] poppler && brew install --cask libreoffice. Windows: install PowerPoint desktop, Node.js, Python 3.12, LibreOffice, and Poppler, and use py -3.12 -m venv .venv for the dependencies. Installation for other hosts is not documented in the README.
How do you use this skill?
Invoke $deep-ppt in a Skills-capable agent with your topic, materials, and target scenario, e.g.: "Use $deep-ppt to create a 12-slide Chinese lab meeting deck from this paper PDF. Keep the deepPPT visual style, include rendered QA, and provide the final PPTX." You can also run scripts/check_env.py manually and, after generation, run normalize_pptx.py, check_pptx.py, extract_pptx_text.py, qa_layout.py, and render_preview.py. SKILL.md defines init/plan/build/check/render/review/complete/resume commands invoked through the scripts/deepppt.py wrapper.