PPT Agent
Generate professional presentations like software engineering: a multi-subagent pipeline with pixel-level visual QA that turns one sentence into a deliverable PPTX.
SKILL.md shows substantial safety design: subagent context isolation, output confined to relative paths, mandatory WAIT_USER confirmation points, forced manual-audit release gates, and prohibition on the main agent editing formal artifacts. Deducted: reliance on web search and Puppeteer with undisclosed dependency/security posture; prompt templates and scripts are absent from evidence so data-flow transparency is only partially confirmable; no explicit overall rollback/cleanup mechanism.
High internal self-consistency at the documentation level: Gate validators, retry/rollback, and checkpoint-recovery rules are complete and coherent. Deducted: all key paths depend on scripts/*.py and references/ files that are not provided, so no command can be statically reproduced; the 'scripts are black boxes, never read them' rule shifts debugging cost to the user on interface mismatch; no test or CI evidence, and the static ceiling of 10 applies.
Target scenario is clear (Chinese-language end-to-end PPT generation), environment sensing and degradation strategies are specified, Chinese support is good. Deducted: the frontmatter trigger description is excessively broad ('beautify', 'visualize data', implicit intents all trigger), creating false-trigger risk; non-fit boundaries are weakly declared; heavy dependence on a CLI that must support subagent creation and model override, sharply reducing availability elsewhere.
Versioning (v4.1 / 2026.04.09), changelog, MIT LICENSE, directory layout, and single-source indices are clear and consistently named. Deducted: repository license metadata is NOASSERTION, inconsistent with LICENSE; publisher unverified with no maintenance commitment or update path; the '2026-04-09' date is anomalous; critical files (cheatsheet, templates) are absent so maintainability cannot be fully confirmed.
Claims to turn a one-line request into directly usable PPTX, which would be high marginal value. Deducted: no statically verifiable output evidence (screenshot assets not included in evidence); the process is very heavy—multiple subagents, many user waits, high token cost—so cost/benefit is questionable; the core value claim rests on author statements, and the static ceiling is below 7.
Versioned changelog, artifact chain, and Gate design provide auditable internal consistency. Deducted: no CI workflow, no committed test suite, no third-party execution evidence; README marketing claims ('zero dropped fonts', 'pixel-level') are unverifiable; showcase images are not in the evidence and fact/marketing are not separated.
- Core scripts and template files were not in the reviewed evidence; none of the Gates or commands have been execution-verified, so real-world usability is unknown.
- Trigger conditions are over-broad; the skill may falsely activate on mere beautification or data-visualization requests and launch a very heavy multi-agent pipeline.
- Depends on web search and Puppeteer rendering; security and network reachability (especially from mainland China) are unverified.
- License metadata is NOASSERTION, inconsistent with the MIT LICENSE file; verify via the install channel.
- Token and time costs are high, and output-quality claims are author statements only; run a small pilot before production use.
What does this skill do, and when should you use it?
PPT Agent (v4.1) is a code-driven, fully automated presentation generation framework that orchestrates multiple subagents via a strict state machine. It splits deck production into seven stages — interview, branch confirmation, research/source compression, outline, style lock, per-page parallel production (Planning → HTML → Visual QA), and dual-format export — with each stage's artifacts validated by gates before proceeding. Pages render HTML only after passing a validated JSON planning contract, then get screenshotted and visually audited, with structural DOM+CSS rewrites to fix overflow, addressing the hallucination, overlap, and layout chaos typical of LLM-generated slides. Final deliverables are a web preview plus PNG and SVG PPTX exports.
Reads user interview answers and local files (or performs web research), generates subagent prompts from templates via prompt_harness.py, and dispatches ResearchSynth / SourceSynth / Outline / Style / PageAgent subagents to produce search-brief.txt, outline.txt, style., planningN., and slide-N.html. It runs two-layer validation with contract_validator, planning_validator, and visual_qa.py; after each page's HTML renders, it screenshots the page and has a subagent rewrite DOM+CSS to eliminate layout conflicts. Finally, a Puppeteer-based rasterization engine exports preview.html, presentation-png.pptx, presentation-svg.pptx, and delivery-manifest. into ppt-output/runs/<RUN_ID>/.
- Professionals who need a structured report for their boss — a single sentence like 'make me a deck about X' triggers the full pipeline
- Founders building an investor pitch deck, e.g. a 15-page dark tech-style roadshow presentation
- Teams with existing doc/excel/pdf/pptx material who want it converted into a deck, or just beautified/restructured
- Trainers generating multi-page courseware who want to personally review key pages via the manual audit mode
- Interrupted tasks: supply the RUN_ID and the framework scans on-disk artifacts to infer the resume point and continue
What are this skill's strengths and limitations?
- Stage-isolated subagents keep contexts uncontaminated; models are explicitly pinned to prevent silent downgrade
- Data/render layer separation: planning JSON passes validation before HTML renders, so bad structure is caught early
- Pixel-level Visual QA loop: screenshots are audited and fixed via DOM+CSS rewrites, not spacing nudges
- Stateless checkpoint recovery: trusts only disk artifacts and gate checks, resumable across conversations
- Dual PPTX export: PNG pipeline for pixel-perfect fidelity, SVG pipeline with editable fonts
- GitHub license field is NOASSERTION (README says MIT) — an inconsistency to verify before adoption
- Heavy runtime requirements: Python, a screenshot engine, and subagent-creation capability; bare model APIs won't run it directly
- The 7-stage pipeline with mandatory interview and multiple gates is overkill for small tasks; dead subagents force full page reruns
- Blocked pages require human adjudication — long runs need someone watching
- No test suite or cross-platform compatibility evidence in the source; output quality depends on the chosen model
How do you install this skill?
Run npx skills add sunbigfly/ppt-agent-skills to install the skill into an Agent Skills-compatible agent environment. The repo contains SKILL.md (main console), scripts/ (validators, harness, exporters), references/ (playbooks, styles, layouts, charts, blocks), and assets/. Ensure the environment has Python 3 and a working screenshot/Puppeteer engine; the research branch also needs web search tooling.
How do you use this skill?
Trigger it with natural language in any Skill-capable agent environment, e.g.: "Create a 15-page pitch deck on embodied intelligence trends for 2026, dark tech style." The flow starts with a mandatory interview (Step 0) covering audience, page density, image policy, subagent model and thinking effort, and whether manual audit is on; then the P0→P5 pipeline runs, writing all outputs to ppt-output/runs/<RUN_ID>/. Exact CLI parameters live in the repo's cli-cheatsheet.md (not enumerated in the source material).