# Ppt Cast > Create or maintain quality-first character-video PowerPoint decks in which one consistent protagonist performs every content slide while text, data, charts, UI, logos, and sources remain native/editable. Use for 角色出演PPT, 每页视频PPT, mascot/character video decks, Deckformance, or /ppt-cast. Do not use for ordinary static decks or full-page AI videos with burned-in text. - Skill: `lhfer/ppt-cast` (Agent Skill, multi-file: 103 files) - Install (CLI): `npx skillmds@latest add lhfer/ppt-cast` - Raw SKILL.md: https://api.skillmd.com/api/skills/lhfer/ppt-cast/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Docs & Writing - Author: lhfer (https://skillmd.com/u/lhfer) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lhfer/ppt-cast --- # Deckformance (`ppt-cast`) Create an editable dynamic PPTX in which the same protagonist performs each content slide's idea. Cover and true section pages may be static; content and closing pages may not. ## Non-negotiable contract - Use v2 for every new job. Preserve v1 readers/builders for existing jobs; never rewrite a v1 job in place. - Plan content, evidence, typography, composition, video slot, motion, and registered assets before generating media. - Identity references determine who the character is, not crop, pose, background, scale, or camera distance. A bust/headshot must first produce a validated full-body performance model. - A dynamic page has exactly one same-protagonist video layer. Never substitute a still, empty slot, unrelated old media, or secondary character when video fails. - Dynamic slide video must be generated by a real external video-generation model. MotionPlan specifies direction and playback; it never synthesizes video bytes. - Video owns action, emotion, and metaphor. Native PPT owns readable text, exact data, charts, tables, UI, logos, annotations, sources, and speaker notes. - Poster/video bytes must already satisfy final aspect, codec, fps, silence, and size before five-frame QA. Any later crop/transcode/compression invalidates downstream evidence. - External facts and structured values must be source-backed and compiled into the corresponding notes under `[Sources]`. - Fail closed on unresolved fonts, unreliable text fit, unsupported provider/renderer/player capabilities, stale hashes, or incomplete receipts. Do not hand-write success booleans. - `candidate.pptx` means automatic package/render/evaluation gates passed. `final.pptx` additionally means the exact same bytes passed real PowerPoint playback. Human visual acceptance remains a separate state. For required dynamic-slide media, fixture/synthetic bytes, `import-video`, local files, static images, native PowerPoint animation, CSS animation, canvas/programmatic motion, or any other non-model substitute are test/draft material only and must not enter `candidate.pptx`. This restriction does not change the `static-native` cover/section allowance or the use of native editable layers for text and evidence. ## Route the task For a new v2 production job, read [v2-production.md](references/v2-production.md). It covers the artifact chain, Typography, semantic composition, Hybrid layers, Motion, media receipts, deterministic deck compilation, candidate build, and v1 migration boundary. Before starting a formal run, integrating provider/render infrastructure, generating host distributions, or attempting candidate/final release, read [runtime-and-release.md](references/runtime-and-release.md). For regression suites, real-generation evaluation, gallery publication, or v1/v2 blind review, read [benchmarks-and-gallery.md](references/benchmarks-and-gallery.md). Benchmark definitions and old jobs are not proof that current acceptance gates passed. For identity/body construction, additionally read [character-model.md](references/character-model.md). For generation/frame/slot inspection details, read [qa.md](references/qa.md). Load the felt-specific human guide only for `felt-yarn` work: [felt-yarn.md](references/presets/felt-yarn.md). ## Focused felt editorial master When the user asks for the reference-led premium felt card composition, choose `felt_editorial_split_master` instead of adding more Composition families. Read [felt-master-template-design-principles.md](references/felt-master-template-design-principles.md) and load `.grok/skills/ppt-cast/scripts/felt_editorial_split_master.js`. The master has one 2.08:1 centered card and two strict mirrors only: `media-left` and `media-right`. Use the selected 57/43 media/panel ratio unless the user is explicitly reviewing the three ratio candidates. Keep the native headline, body, and closing line in the charcoal panel; generate the media at a near-6:5 aspect and keep its subject inside the central safe area. “Clean” does not mean backgroundless: final media needs a story-specific felt miniature environment with visible foreground, midground, and background depth, not a generic studio wall or empty tabletop. Supported page inputs are intentionally narrow: `mediaSide`, `eyebrowOrIcon`, `pageNumber`, `headline`, `body`, `closingLine`, `videoPoster`, `videoSlot`, and optional `characterSet`. `characterSet.referenceImages` accepts one or two job-relative user uploads; there is no default mascot. Preserve the uploaded identity while recomposing the felt miniature scene, and never treat sample characters as template identity. Before rendering, copy the packaged `.grok/skills/ppt-cast/assets/felt-edge-negative-mask.png` to `/assets/felt-edge-negative-mask.png`. Apply it as the card-level negative alpha overlay so the warm canvas color remains adaptive around both panel and media. Build the charcoal panel from `model.panelShape` (or `panelShapeConfigForFeltMaster`) rather than reconstructing its stroke: it uses a true no-fill outline (`fill: "none"`, `width: 0`). Never give the panel a zero-width solid charcoal line; PowerPoint exposes that as a one-pixel hairline outside the mask. As a final package gate, inspect every `felt-panel-*` in the exported PPTX and require `` with no solid line fill. Do not replace the mask with a sawtooth shape, generic scallop, or two independently rounded rectangles. ## v2 default workflow 1. Run `doctor`, initialize the v2 job, and author approved content/visual contracts. Candidate does not require PowerPoint; final does. 2. Compile `design-plan.json` before media generation. Use `preview_design_v2.js` when an early native-layer/typography preview is useful; its video slots are placeholders and it is never releasable. 3. Route the host to a concrete external video-generation adapter, invoke the real model, and produce exact final MP4 bytes plus a hash-bound `generate-video` provider receipt. Then produce media-budget, five-frame, slot, and evaluation receipts. If the provider, final MP4, or receipt is missing, stop before `media-ready`; never fall back to import or synthetic motion. 4. Compile deterministic `deck.json`, build only `candidate.staging.pptx`, then run the independent v2 package validator. 5. True-render the exact staging bytes, inspect every slide, and bind the renderer implementation/hash and page outputs. 6. Run release `preflight`. Treat `ready`, `degraded`, and `blocked` literally, then let `jobctl_v2` publish `candidate.pptx` only after all automatic gates pass. 7. On a PowerPoint-capable Mac, play and capture the exact candidate bytes. The builder/package gate proves timing XML was authored; only this real-player observation proves autoplay. Only the external exact-byte receipt may authorize `jobctl_v2` to publish `final.pptx`. Host adapters share one contract: Grok maps its real external video-generation tool/API to `ProviderAdapter.generateVideo`; Qwen maps a configured MCP or long-running generation service to that operation; a generic host registers an API/MCP/CLI-backed provider adapter. Merely declaring a host/provider name or supplying existing bytes is not a successful adapter invocation. Core commands: ```bash node /scripts/doctor.js --release candidate --json node /scripts/jobctl_v2.js init jobs/ --job-id node /scripts/compile_design.js jobs/ node /scripts/compile_deck_v2.js jobs/ node /scripts/build_deck_v2.js \ jobs/ jobs//build/candidate.staging.pptx \ --release candidate --report jobs//qa/build-report.json node /scripts/validate_pptx_v2.js \ jobs/ build/candidate.staging.pptx \ --release candidate --report qa/package-qa.json ``` Then follow the true-render and `jobctl_v2` commands in [v2-production.md](references/v2-production.md). Do not skip directly from a successful build to a published candidate. ## v1 compatibility only For an existing v1 job, use the legacy chain documented in [contracts.md](references/contracts.md): `jobctl.js`, `compile_deck.js`, `build_deck.js`, `validate_pptx.js`, and `validate_job.js`. Legacy jobs and fixtures remain useful regressions but are not v2 gallery successes. `migrate_v1.js` creates a blocked v2 draft requiring explicit redesign. Migration alone never authorizes design compilation, candidate release, or gallery eligibility. ## Claim discipline Keep these states separate in reports: - automated tests passed; - real external video-generation model invoked with bound provider/model/request/output evidence; - trusted true render passed; - candidate released for exact bytes; - PowerPoint playback passed for exact bytes; - human visual/blind review accepted. If no concrete external video provider is callable, stop at the design/preview stage and say so; a dynamic-slide candidate is not permitted. If this Mac has no PowerPoint, stop at candidate and say so. Never describe fixture/synthetic/import media as model-generated, package timing as real-player autoplay, or unexecuted fixed/real benchmarks, gallery generation, five-reviewer blind testing, or final-macOS playback as completed.