Research Deck PPTX Skill
Turns structured research and consulting content into dense, institutional-grade PPTX decks with strict content fidelity and step-by-step confirmation gates.
SKILL.md shows good safety design: explicit user confirmation gates at every step (G1), content fidelity rules prohibiting fabricated data (G3), state persistence enabling rollback after interruption (G2 and Resume Protocol), and explicit warnings when non-editable elements are used (G5). However, it is not demonstrably least-privilege: it depends on scripts and workflows under ~/.claude/skills/pptx that are not present in this repository and cannot be verified statically; no standalone sensitive-data handling guidance. Deductions: unobservable external dependencies, missing data-flow disclosure.
Instructions are self-consistent, with a clear state machine and resume protocol and rollback paths on failure. But hardcoded paths (~/.claude/skills/pptx/scripts/...) and the QA protocol of an external pptx skill mean key paths are not reproducible if those files are missing or relocated; error handling for abnormal input defers to external docs with no skill-specific guidance; no test evidence. Below the static cap of 10.
Trigger words are explicit (including Chinese triggers), the frontmatter declares non-fit boundaries (no general presentations, no editing existing PPTX, no pitch decks), and inputs/expected outputs are described. However, environment fit depends on a full local toolchain (Node.js, LibreOffice, poppler, markitdown) with no installation verification or environment checks; if the base pptx skill is unreachable the skill is essentially unusable, and this dependency risk is not disclosed.
Information architecture is well layered (progressive disclosure: only the current step's reference file is read), and the README provides installation, file structure, and design principles. But there is no version number/changelog, no maintenance ownership or update path; license metadata is unknown while the README claims MIT, and the reference files were not present in the reviewed material so hidden assumptions cannot be checked.
The claimed value — high-density, editable research decks — is clear with plausible marginal value over manual creation, and design rules are concrete. But there are no generated examples, screenshots, or verified outputs, so direct usability is unproven; a higher score is unreachable under static review.
Only author-authored design documentation is present: no test suite, no CI evidence, no third-party execution records, and no sample outputs to inspect. Facts and inference are reasonably separable (rules are explicit), but key claims (QA loop effectiveness, rendering quality) are entirely unverifiable independently.
- The skill hard-depends on scripts and docs under ~/.claude/skills/pptx; it will not run without the companion pptx skill installed — verify dependencies first.
- Requires a local toolchain (Node.js, LibreOffice, poppler, markitdown, pdfplumber), and Step 3b fact-checking depends on web search capability.
- License metadata is unknown; the README only verbally claims MIT — verify legal terms before use.
- No versioning or changelog, and the reference files were not provided in the reviewed material, so actual step content is unverified.
- No sample outputs or test evidence; output quality must be validated by the user.
What does this skill do, and when should you use it?
This is a Claude Code skill for generating high-density, chart-heavy PPTX decks for investment research and consulting scenarios. It runs a six-step pipeline — content input, format selection, outline, per-page storyline, logic and fact-check review, chapter-by-chapter rendering — with an explicit user confirmation gate at every transition. Content fidelity is strictly enforced: no invented data, and content gaps are flagged rather than silently filled. It reuses the base pptx skill's rendering and QA infrastructure, and persists progress to a _deck_state. file so long-running deck builds can resume after interruption.
Reads structured user content (e.g., docx reports, PDFs) and first assesses content sufficiency. At each step it builds the chapter outline, writes conclusion-oriented slide titles and key messages, and fact-checks the storyline (with web search). For rendering it chooses between XML editing (user-provided template) or PptxGenJS from scratch, maximizing native editable charts, then runs content QA (markitdown) and visual QA (soffice → PDF → image inspection) after each chapter. After every step it writes progress to _deck_state., and on restart it detects the state file and offers to resume, roll back to a step, or start over.
- A sell-side or buy-side analyst converting a research memo into a chart-dense equity research deck for internal review
- A consultant producing an industry analysis deliverable deck from completed analysis work
- An investment team generating equity research presentation slides where every chart must be paired with analytical commentary, not big hero numbers
- An industry researcher who needs storyline logic and facts verified before layout and rendering
- A user on a long deck build who wants to interrupt and resume exactly where they left off
What are this skill's strengths and limitations?
- Forced user confirmation at every step and chapter keeps the build controllable and correctable
- Strict content fidelity — no invented data; gaps are explicitly flagged
- State persistence via _deck_state. enables resume and rollback to any step
- Maximizes native editable charts, so outputs remain editable in PowerPoint
- Logic review (step 3b) happens before visual design, avoiding rework
- Hard dependency on the base pptx skill (~/.claude/skills/pptx/), which must be installed first and whose setup is not documented here
- Heavy dependency chain: Node.js, LibreOffice, poppler, and multiple Python packages are all required
- Fact-checking (step 3b) requires web search capability, so it cannot run fully offline
- Repository has no description or topics; license is unknown in metadata (README claims MIT)
- No test suite and no rendered sample decks or screenshots as quality evidence
- Explicitly unsuitable for general presentation creation, editing existing PPTX files, or pitch/roadshow decks
How do you install this skill?
Clone the repo into your skills directory:
git clone https://github.com/zhangxinrui127-stack/pptx-research.git ~/.claude/skills/pptx-research
Prerequisites: Node.js, LibreOffice (soffice), pdftoppm (poppler), and Python packages pdfplumber and markitdown. The skill reuses the base pptx skill at ~/.claude/skills/pptx/; installing that base skill is an implicit prerequisite whose installation steps are not documented here.
How do you use this skill?
In Claude Code, invoke with trigger phrases like "Create a research deck from this report", "Generate equity research slides", or "Build a consulting deliverable deck" (Chinese triggers include 研报PPT, 咨询报告PPT, 行研PPT). The skill will ask for confirmation at each stage: content sufficiency → language/template → outline → per-page storyline → logic/fact review → per-page visuals → chapter rendering, and you must explicitly approve each one. After an interruption it detects _deck_state. and asks where to resume from.
How does this skill compare with similar options?
The skill itself draws a clear line against the base pptx skill: the pptx skill covers general deck creation and editing of existing files, while this skill targets high-density informative research/consulting decks — they are complementary, not competing.