From cc9e1fd94ec5498b0adeff54a71df359a0ea0388 Mon Sep 17 00:00:00 2001 From: Haitao Pan Date: Tue, 2 Jun 2026 07:11:29 +0800 Subject: [PATCH] feat(skills): add workflow orchestration skills --- README.md | 5 ++ skills/README.md | 1 + .../article-to-image-series/SKILL.md | 73 ++++++++++++++++ .../content-pdf-with-images/SKILL.md | 84 +++++++++++++++++++ .../workflows/image-series-to-video/SKILL.md | 74 ++++++++++++++++ .../workflows/research-to-slide-deck/SKILL.md | 72 ++++++++++++++++ 6 files changed, 309 insertions(+) create mode 100644 skills/workflows/article-to-image-series/SKILL.md create mode 100644 skills/workflows/content-pdf-with-images/SKILL.md create mode 100644 skills/workflows/image-series-to-video/SKILL.md create mode 100644 skills/workflows/research-to-slide-deck/SKILL.md diff --git a/README.md b/README.md index dfc0a5b..e090366 100644 --- a/README.md +++ b/README.md @@ -9,6 +9,7 @@ - `skills/image-production/`：连续风格 PNG / 信息图素材生成。 - `skills/audio-production/`：视频音效、BGM 和音频素材工作流。 - `skills/animation/`：HyperFrames 动效、简笔画动画和 Anime.js 适配。 +- `skills/workflows/`：跨技能编排工作流，定义阶段、下游产物、验收门和交付合同。 - `skills/workspace-core/`：从 OpenClaw/Codex runtime 同步回来的 workspace core skills。 其中 source-owned 内容生产技能基于 `HyperFrames` 工作流，目标是把“想法 -> 脚本 -> 画面 -> 音频 -> 成片”的流程沉淀成可复用方法。Workspace core skills 只作为 runtime 辅助技能归档，不与内容生产技能混放。 @@ -32,6 +33,10 @@ | 视频音效工作流 | 音效搜索、下载与合成、时间线接入 | `skills/audio-production/sound-fx-for-video/SKILL.md` | | 简笔画动画视频 | 线稿风 + 短画面字；**主动网络搜参考图临摹**；逼真非抽象；GSAP 主时间线 + 可选 Anime.js；抽检闭环 | `skills/animation/sketch-animation-video/SKILL.md` | | Anime.js（HyperFrames） | seek 驱动适配、`window.__hfAnime` 注册、与 GSAP 分工 | `skills/animation/animejs/SKILL.md` | +| 图文 PDF 工作流 | 主题拆章节、生成章节图、汇总排版为 PDF | `skills/workflows/content-pdf-with-images/SKILL.md` | +| 文章转图片系列 | 将文章、提纲或主题序列转成 manifest-backed 多图系列 | `skills/workflows/article-to-image-series/SKILL.md` | +| 图片系列转视频 | 从 PNG manifest 生成讲解视频并执行验收 | `skills/workflows/image-series-to-video/SKILL.md` | +| 研究到幻灯片 | 将研究资料、来源和主题 brief 转成可验证 slide deck | `skills/workflows/research-to-slide-deck/SKILL.md` | ## Workspace Core Skills diff --git a/skills/README.md b/skills/README.md index 859c9b1..3eda9d6 100644 --- a/skills/README.md +++ b/skills/README.md @@ -9,6 +9,7 @@ | `image-production/` | 连续风格 PNG / 信息图素材 | | `audio-production/` | SFX、BGM、音频素材工作流 | | `animation/` | HyperFrames 动效、简笔画动画、Anime.js 适配 | +| `workflows/` | 跨技能编排工作流、阶段验收、交付合同 | | `workspace-core/` | OpenClaw/Codex runtime core skills | Source-owned 内容生产技能和 runtime 同步技能使用目录边界隔离，不再使用并列顶层 `workspace-core-skills/`。 diff --git a/skills/workflows/article-to-image-series/SKILL.md b/skills/workflows/article-to-image-series/SKILL.md new file mode 100644 index 0000000..0d982cc --- /dev/null +++ b/skills/workflows/article-to-image-series/SKILL.md @@ -0,0 +1,73 @@ +--- +name: article-to-image-series +description: "文章转图片系列工作流：将文章、提纲或主题序列转成风格一致的多图系列。Workflow skill for converting an article, outline, or topic sequence into a coherent multi-image series; use for one image per chapter, article illustrations, infographic series, Xiaohongshu-style cards, or manifest-backed image sets." +--- + +# Article To Image Series Workflow + +This workflow turns text structure into a validated image series. It is intentionally staged so the agent plans image count and prompts before generating assets. + +## Output Contract + +Final deliverables must stay inside the current task artifact scope: + +- `workflow.plan.md` +- `series.config.json` +- `prompts/image-prompts.md` +- `assets/images/*.png` +- `assets/images/manifest.md` +- `DELIVERY.md` + +## Phase 1: Parse Source + +Read the user input and identify: + +- topic or article title +- narrative line +- target audience +- chapter or card count +- style requirements +- any required reference images or source URLs + +If no count is given, choose a compact 3-7 image series based on the narrative. + +## Phase 2: Series Plan + +Write `workflow.plan.md` and `series.config.json`. + +Each image entry must define: + +- `chapter_id` +- `title` +- `subtitle` +- core message +- visual metaphor +- key labels +- output path under `assets/images/` + +Do not ask an image model for a batch collage. One entry means one standalone output file. + +## Phase 3: Prompt Pack + +Write `prompts/image-prompts.md` with one prompt per image. + +Prompts must keep a consistent family style while preserving per-chapter meaning. Avoid generic cyberpunk, stock-style, and unreadable text-heavy visuals. + +## Phase 4: Generate Images + +Generate or edit images one by one. Save each final PNG under `assets/images/`. + +If an image tool writes to a cache directory, copy the final image into the artifact scope. Do not leave final outputs in `Downloads`, model cache, `/tmp`, or global media folders. + +## Phase 5: Manifest And Validation + +Write `assets/images/manifest.md` with one row per image. + +Required checks: + +- PNG count equals manifest row count. +- Every manifest `file` exists and is a real PNG. +- No manifest file path points outside the artifact scope. +- No generated image is a combined grid, contact sheet, or storyboard unless explicitly requested. + +Stop on validation failure and report the missing or invalid image file. diff --git a/skills/workflows/content-pdf-with-images/SKILL.md b/skills/workflows/content-pdf-with-images/SKILL.md new file mode 100644 index 0000000..f707005 --- /dev/null +++ b/skills/workflows/content-pdf-with-images/SKILL.md @@ -0,0 +1,84 @@ +--- +name: content-pdf-with-images +description: "图文 PDF 工作流：将主题或文章拆成章节，为每章生成配图，并汇总排版为真实 PDF。Workflow skill for turning a topic or article request into a chaptered PDF with one image per chapter; use for illustrated PDF reports, chapter images, and final PDFs assembled from text plus generated images." +--- + +# Content PDF With Images Workflow + +This is an orchestration skill. It does not replace planning, image, or PDF skills. It forces a staged handoff so the final artifact is a real PDF with real image inputs, not a placeholder file. + +## Output Contract + +Final deliverables must stay inside the current XWorkmate/OpenClaw task artifact scope: + +- `workflow.plan.md` +- `article.md` +- `assets/images/*.png` +- `assets/images/manifest.md` +- `prompts/image-prompts.md` +- `series.config.json` +- `exports/final.pdf` +- `DELIVERY.md` + +## Phase 1: Plan + +Create `workflow.plan.md` before writing final content. It must include: + +- topic and audience +- chapter list +- target word or character count +- one image concept per chapter +- downstream skill sequence +- final acceptance checklist + +If the user gives an explicit chapter sequence, preserve it literally. + +## Phase 2: Article + +Write `article.md` with: + +- title +- one section per planned chapter +- concise intro and closing summary +- total length matching the user request + +Do not claim completion after this phase. + +## Phase 3: Images + +Use the image-series skill path for one standalone PNG per chapter. + +Required checks before continuing: + +- PNG count equals chapter count. +- Every `assets/images/*.png` is a real non-empty PNG. +- `assets/images/manifest.md` exists and references only relative paths inside the artifact scope. +- `prompts/image-prompts.md` and `series.config.json` exist. + +If any check fails, stop and report the missing image artifact. Do not generate a placeholder PDF. + +## Phase 4: PDF + +Use the PDF creation path to assemble `article.md` plus image manifest into `exports/final.pdf`. + +The PDF must contain: + +- title page +- chapter text +- matching chapter image near each section +- closing summary + +Use a PDF library such as reportlab when available. The file must be large enough to plausibly contain text and images; a tiny one-page placeholder PDF is a failure. + +## Phase 5: Delivery Check + +Before final response, run equivalent checks: + +```bash +find . -maxdepth 4 -type f | sort +file exports/final.pdf +pdfinfo exports/final.pdf +pdftotext exports/final.pdf - | head -40 +``` + +Only report success if `exports/final.pdf` exists and the expected source files are present. diff --git a/skills/workflows/image-series-to-video/SKILL.md b/skills/workflows/image-series-to-video/SKILL.md new file mode 100644 index 0000000..c077d56 --- /dev/null +++ b/skills/workflows/image-series-to-video/SKILL.md @@ -0,0 +1,74 @@ +--- +name: image-series-to-video +description: "图片系列转视频工作流：将带 manifest 的 PNG 图片系列转成带讲解、字幕和验收的 MP4 视频。Workflow skill for converting a manifest-backed image series into a narrated video; use when generated PNGs or an image manifest should become an MP4, long-image explanation video, chapter video, narrated walkthrough, or social video." +--- + +# Image Series To Video Workflow + +This workflow starts only after a real image series exists. It prevents the video stage from inventing fake screenshots, CSS cards, or placeholder images. + +## Required Inputs + +The current artifact scope must contain: + +- `assets/images/*.png` +- `assets/images/manifest.md` + +The manifest must describe one row per image with title, file, usage, scan mode, and focus guidance. + +## Output Contract + +Final deliverables must stay inside the current task artifact scope: + +- `video.config.json` +- `index.html` +- `assets/audio/*.mp3` +- `assets/audio/bgm.wav` +- `renders/*.mp4` +- `ffprobe.json` +- `DELIVERY.md` + +## Phase 1: Input Gate + +Validate before building video: + +```bash +find assets/images -maxdepth 1 -type f -name '*.png' | sort +test -f assets/images/manifest.md +``` + +Reject the task if images are missing, empty, non-PNG, or not referenced by manifest. + +## Phase 2: Video Plan + +Create or update `video.config.json` from the manifest: + +- one scene per image or chapter +- title and caption per scene +- scan mode and safe focus from manifest +- voiceover text +- inspect timestamps +- output file name + +Do not hand-write fixed scene arrays unrelated to the manifest. + +## Phase 3: Build + +Use the repository video runner when available. Otherwise create the minimal equivalent pipeline: + +- generate `index.html` +- generate voiceover audio +- run visual acceptance snapshots +- render MP4 +- run ffprobe + +## Phase 4: Acceptance + +Only report completion if: + +- `renders/*.mp4` exists and is non-empty. +- `ffprobe.json` shows video and audio streams. +- duration is plausible for the number of scenes. +- sampled frames show real source images, visible captions, and no obvious overlap. + +If any step fails, report the failing phase and keep partial artifacts for diagnosis. diff --git a/skills/workflows/research-to-slide-deck/SKILL.md b/skills/workflows/research-to-slide-deck/SKILL.md new file mode 100644 index 0000000..6f6629b --- /dev/null +++ b/skills/workflows/research-to-slide-deck/SKILL.md @@ -0,0 +1,72 @@ +--- +name: research-to-slide-deck +description: "研究到幻灯片工作流：将研究资料、来源和主题 brief 转成可验证的演示文稿。Workflow skill for turning research notes, sources, or a topic brief into a verified slide deck; use for PPT, PPTX, slides, research presentations, briefing decks, or source-backed presentations." +--- + +# Research To Slide Deck Workflow + +This workflow converts research into a presentation with explicit source, outline, slide, and verification phases. + +## Output Contract + +Final deliverables must stay inside the current task artifact scope: + +- `research/sources.md` +- `deck-outline.md` +- `slides/slide-notes.md` +- `exports/deck.pptx` or `exports/deck.pdf` +- `DELIVERY.md` + +When a PPTX writer is unavailable, produce a PDF deck and state that format clearly. + +## Phase 1: Research + +Gather or parse source material into `research/sources.md`. + +Each source note should capture: + +- title +- URL or local path when available +- date if relevant +- key facts +- how it supports the deck + +For current facts, live sources, pricing, market data, product specs, or recent events, verify against live sources before writing the deck. + +## Phase 2: Deck Outline + +Write `deck-outline.md` before creating slides. + +The outline must include: + +- audience and purpose +- 6-12 slide titles unless the user specifies otherwise +- one message per slide +- required charts, screenshots, tables, or diagrams +- source mapping per slide + +Do not begin slide generation until the outline is coherent. + +## Phase 3: Slide Production + +Create the deck from the outline: + +- title slide +- agenda or context slide when useful +- evidence slides with concise headings +- visual slides for comparisons, timelines, architecture, or workflow +- closing recommendation or next steps + +Use charts, tables, diagrams, or screenshots only when they support the message. Avoid decorative filler. + +## Phase 4: Verification + +Before final response: + +- confirm `exports/` contains the final deck +- render or inspect the deck when tooling is available +- check slide count +- check that source-backed claims map to `research/sources.md` +- check text does not overflow obvious slide containers + +If verification cannot run, report the missing verifier and keep the generated deck plus source files.