Communitygithub.com

NeverSight/learn-skills.dev

自动从智能体轨迹、日志、侧车工件和重复的用户偏好中维护个性化回写。当任务产生可重用的偏好、经验教训、私有用户记忆、项目契约或潜在的公共技能规则时使用,无需中断用户。

learn-skills.dev 是什么?

该技能充当智能体的个性化记忆写入层,使其能够随着时间的推移学习用户行为并优化交互,而无需为每个偏好更新打断用户。它支持多种平台,包括Claude Code、Codex和兼容的AI编程助手。核心机制是通过低成本侧车扫描从轨迹、日志、差异和侧车输出中提取洞察,将观察结果分类为私有用户偏好、项目契约或公共技能规则候选,并应用置信度门控来决定存储级别。该技能确保不存储原始隐私数据,防止 collaborator 特定信息泄露,并通过渐进式加载策略引用详细指南和模板,最终将可重用的经验教训嵌入到持久化存储中,例如用户工作空间、项目记忆文件或公共技能规则库。

兼容平台✓Claude Code✓Codex CLI~Cursor
npx skills add https://github.com/NeverSight/learn-skills.dev/tree/HEAD/data/skills-md/a-green-hand-jack/ml-research-skills/personalization-memory

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

个性化记忆

将重复的交互痕迹转化为持久的偏好,而无需让主智能体停下来询问用户每个记忆更新。此技能是“系统随时间变得更个性化”的回写层。

技能目录布局

<installed-skill-dir>/
├── SKILL.md
├── references/
│   ├── trajectory-scanner.md
│   └── writeback-policy.md
└── templates/
    └── preference-ledger.md

核心契约

  • 不要仅仅为了决定是否记住常规偏好而询问用户。
  • 优先使用低成本侧车扫描,从日志、轨迹、差异、侧车输出、审查包和项目记忆中综合信息。
  • 切勿将原始私人对话、原始日志、凭据、仅本地路径或协作者消息复制到公共仓库记忆中。
  • 存储最小的可重用经验教训,而不是整个事件。
  • 将个性化视为一个晋升阶梯:观察 -> 候选 -> 偏好 -> 项目契约 -> 可重用技能规则候选。
  • 公共技能的变更需要与其他技能变更相同的标准仓库验证、提交、推送和重新安装流程。

使用时机

在以下情况下使用此技能:

  • 用户表示工作流应个性化、自动化、被记住或从轨迹中学习
  • 智能体注意到重复的偏好,例如“不要问我”、“先提交/推送”、“使用 Overleaf 编译”、“偏好截图/页面捆绑”、“使用侧车进行低风险扫描”
  • 其他技能完成了有意义的工作,并且存在可重用的工作流、写作、图形、LaTeX、Git、审查或计算经验教训
  • 应扫描 Codex 或 Claude Code 轨迹日志以生成候选偏好
  • 项目记忆需要将私有用户事实与共享项目契约和公共技能规则分离

与 sidecar-task-runner 配对用于低成本扫描,当项目级结论被接受并需要更新 memory/ 时,与 research-project-memory 配对。

渐进式加载

  • 在决定偏好归属之前,阅读 references/writeback-policy.md。
  • 在启动对日志、轨迹、侧车工件或仓库历史的侧车扫描之前,阅读 references/trajectory-scanner.md。
  • 当项目缺少偏好账本时,使用 templates/preference-ledger.md。

工作流

  1. 收集工件输入。 优先使用经过清理的工件:memory/、最近的 git diff、.agent/sidecars/*/decision.md、.agent/code-reviews/*/fix-log.md、.agent/layout-issues/*/manifest.md、论文/代码的 .agent/ 状态,以及当前任务摘要中用户明确陈述的偏好。
  2. 在有用时运行低成本扫描。 对于非平凡的轨迹扫描,使用 sidecar-task-runner 及其 personalization-scanner 预设。扫描器输出候选结果;它不得直接编辑记忆。
  3. 对每个候选进行分类。 分配范围:private-user、project、public-skill-candidate 或 discard。分配类型:workflow、writing、layout、figure-style、code-review、git、compute、toolchain 或 collaboration。
  4. 应用置信度门控。 观察到一次事件可以成为经验教训。重复出现的事件或用户明确表述可以成为偏好。项目契约需要特定项目的证据。公共技能规则需要维护者意图或跨项目的重复证据。
  5. 写入正确的层级。 遵循 references/writeback-policy.md 的路由规则。编写简短的条目,包含日期、来源工件、置信度和目标范围。
  6. 仅报告有用的摘要。 告知用户记住了什么或哪些候选被推迟。除非下一步会发布私有或面向公众的策略,否则不要请求确认。

默认目标

  • 私有用户偏好:当前智能体的私有记忆区域,例如 ~/.codex/memories/,特别适用于工作站事实、工具别名、本地路径、偏好的交互风格和个人工作流默认值。
  • 项目偏好:memory/、paper/.agent/、code/.agent/ 或 slides/.agent/,特别适用于项目所有智能体共享的契约。
  • 本地扫描工件:.agent/sidecars/<task-id>/ 或 .agent/personalization/<run-id>/;除非经过清理并有意提交,否则保持不受版本控制。
  • 公共可重用技能规则:仅在进行明确的技能维护任务时,写入 skills/<skill-name>/SKILL.md 或 skills/<skill-name>/references/。

候选记录

使用简洁的记录:

- 偏好:<一个可重用的行为>
- 范围:private-user | project | public-skill-candidate | discard
- 类型:workflow | writing | layout | figure-style | code-review | git | compute | toolchain | collaboration
- 证据:<工件路径或用户陈述摘要>
- 置信度:observed | repeated | user-stated | inferred
- 目标:<记忆文件或技能文件>
- 操作:write | defer | promote | reject

安全规则

  • 当衍生规则足够时,不要存储私有的原始轨迹文本。
  • 不要将协作者特定的评论、语音转录、截图或本地文件路径提升为公共技能。
  • 未经主智能体审查,不要让侧车写入最终的公共记忆。
  • 如果候选与现有记忆冲突,将其标记为 candidate 并留下简短注释,而不是覆盖旧规则。
  • 如果偏好会改变外部行为(例如推送、标记、发布、提交作业或发布议题),记录该偏好但保留现有的人工确认环节。

Individual skills in this repo

This repo contains 20 individual skills — each has its own dedicated page.

NeverSight/learn-skills.dev

Use when generating or modifying Remotion video code, creating demo videos, or working with the demo-video/ directory

NeverSight/learn-skills.dev

Create AI avatar and talking head videos via inference.sh CLI. Recommended: P-Video-Avatar (fastest, cheapest, built-in TTS). Also: OmniHuman, Fabric, PixVerse. Audio: Inworld TTS-2 (100+ languages, emotion steering for characters), ElevenLabs, Kokoro. Capabilities: audio-driven avatars, text-to-avatar, lipsync videos, talking head generation, virtual presenters, UGC content. Use for: AI presenters, explainer videos, virtual influencers, dubbing, marketing videos, UGC ads, gaming avatars, NPC dialogue. Triggers: ai avatar, talking head, lipsync, avatar video, virtual presenter, ai spokesperson, audio driven video, heygen alternative, synthesia alternative, talking avatar, lip sync, video avatar, ai presenter, digital human, ugc, ugc video, ugc ad, avatar ugc

NeverSight/learn-skills.dev

Create AI marketing videos for ads, promos, product launches, and brand content. Models: Veo, Seedance, Wan, FLUX for visuals, Kokoro for voiceover. Types: product demos, testimonials, explainers, social ads, brand videos. Use for: Facebook ads, YouTube ads, product launches, brand awareness. Triggers: marketing video, ad video, promo video, commercial, brand video, product video, explainer video, ad creative, video ad, facebook ad video, youtube ad, instagram ad, tiktok ad, promotional video, launch video

NeverSight/learn-skills.dev

Generate AI videos with Google Veo, Seedance 2.0, HappyHorse, Wan, Grok and 40+ models via inference.sh CLI. Models: Veo 3.1, Veo 3, Seedance 2.0, HappyHorse 1.0, Wan 2.5, Grok Imagine Video, OmniHuman, Fabric, HunyuanVideo. Capabilities: text-to-video, image-to-video, reference-to-video, video editing, lipsync, avatar animation, video upscaling, foley sound. Use for: social media videos, marketing content, explainer videos, product demos, AI avatars. Triggers: video generation, ai video, text to video, image to video, veo, animate image, video from image, ai animation, video generator, generate video, t2v, i2v, ai video maker, create video with ai, runway alternative, pika alternative, sora alternative, kling alternative, seedance, happyhorse

NeverSight/learn-skills.dev

ElevenLabs automatic dubbing - translate and dub audio/video into 29 languages while preserving speaker voice via inference.sh CLI. Capabilities: auto speaker detection, voice-preserving translation, video dubbing, audio localization. Use for: content localization, video translation, multilingual content, international distribution. Triggers: dubbing, dub video, translate audio, video translation, audio translation, localize content, elevenlabs dubbing, eleven labs dub, multilingual dub, voice translation, auto dub, language dub, content localization

NeverSight/learn-skills.dev

Explainer video production guide: scripting, voiceover, visuals, and assembly. Covers script formulas, pacing rules, scene planning, and multi-tool pipelines. Use for: product demos, how-it-works videos, onboarding videos, social explainers. Triggers: explainer video, how to make explainer, product video, demo video, video production, video script, animated explainer, product demo video, tutorial video, onboarding video, walkthrough video, video pipeline

NeverSight/learn-skills.dev

Still-to-video conversion guide: model selection, motion prompting, and camera movement. Covers Wan 2.5 i2v, Seedance, Fabric, Grok Video with when to use each. Use for: animating images, creating video from stills, adding motion, product animations. Triggers: image to video, i2v, animate image, still to video, add motion to image, image animation, photo to video, animate still, wan i2v, image2video, bring image to life, animate photo, motion from image

NeverSight/learn-skills.dev

Generate talking head avatar videos with Pruna P-Video-Avatar via inference.sh CLI. Turn a portrait image into a realistic speaking video with built-in TTS. 18x faster and 6x cheaper than competitors. Models: P-Video-Avatar, P-Image (for portrait generation). Capabilities: text-to-avatar, audio-driven avatars, 30 voices, 10 languages, 720p/1080p, built-in TTS, dynamic backgrounds, full-body control. Use for: AI presenters, product demos, explainer videos, virtual influencers, marketing, education, multilingual content, UGC, gaming avatars. Triggers: avatar video, talking head, ai avatar, p-video-avatar, pruna avatar, video avatar, ai presenter, digital human, virtual presenter, lipsync, talking avatar, ai spokesperson, heygen alternative, synthesia alternative, veed alternative, fabric alternative, omnihuman alternative

NeverSight/learn-skills.dev

Generate videos with Pruna P-Video and WAN models via inference.sh CLI. Models: P-Video, WAN-T2V, WAN-I2V. Capabilities: text-to-video, image-to-video, audio support, 720p/1080p, fast inference. Pruna optimizes models for speed without quality loss. Triggers: pruna video, p-video, pruna ai video, fast video generation, optimized video, wan t2v, wan i2v, economic video generation, cheap video generation, pruna text to video, pruna image to video

NeverSight/learn-skills.dev

Render videos from React/Remotion component code via inference.sh. Pass TSX code, get MP4. Supports all Remotion APIs: useCurrentFrame, useVideoConfig, spring, interpolate, AbsoluteFill, Sequence. Configurable resolution, FPS, duration, codec. Use for: programmatic video generation, animated graphics, motion design, data-driven videos, React animations to video. Triggers: remotion, render video from code, tsx to video, react video, programmatic video, remotion render, code to video, animated video, motion graphics code, react animation video

NeverSight/learn-skills.dev

Video ad creation with exact platform-specific specs for TikTok, Instagram, YouTube, Facebook, LinkedIn. Covers dimensions, duration limits, AIDA framework, and caption requirements. Use for: video ads, social media ads, paid media creative, video marketing, ad production. Triggers: video ad, social media ad, tiktok ad, instagram ad, youtube ad, facebook ad, linkedin ad, video creative, ad specs, paid media, video marketing, ad production, reels ad, stories ad, pre roll, bumper ad

NeverSight/learn-skills.dev

Best practices and techniques for writing effective AI video generation prompts. Covers: Veo, Seedance, Wan, Grok, Kling, Runway, Pika, Sora prompting strategies. Learn: shot types, camera movements, lighting, pacing, style keywords, negative prompts. Use for: improving video quality, getting consistent results, professional video prompts. Triggers: video prompt, how to prompt video, veo prompts, video generation tips, better ai video, video prompt engineering, video prompt guide, video prompt template, ai video tips, video prompt best practices, video prompt examples, cinematography prompts

NeverSight/learn-skills.dev

YouTube thumbnail design with specific dimensions, contrast rules, and mobile preview optimization. Covers safe zones, text placement, face expression psychology, and A/B testing. Use for: YouTube thumbnails, video cover images, click-through optimization. Triggers: youtube thumbnail, thumbnail design, video thumbnail, click through rate, ctr optimization, youtube cover, video cover image, thumbnail maker, thumbnail tips, youtube design, video preview image

NeverSight/learn-skills.dev

Landing page conversion optimization with layout rules, hero section design, and CTA psychology. Covers above-the-fold formula, social proof placement, mobile design, and F-pattern reading. Use for: startup landing pages, product pages, SaaS marketing, conversion optimization. Triggers: landing page, hero section, above the fold, conversion optimization, landing page design, cta button, hero image, landing page layout, saas landing page, product page design, conversion rate, landing page best practices

NeverSight/learn-skills.dev

Configure and use the hosted YouTube Data MCP end-to-end with minimal user input. Use when users want the agent to verify Node.js and `npx`, configure MCP server config (Windows/macOS, Cursor/Codex/OpenClaw/OpenCode), request API key at setup time, run post-install capability discovery (`tools/list` and `get_patch_notes`), and then strongly recommend helper skill and Python setup for full local document and spreadsheet workflows.

NeverSight/learn-skills.dev

Creates 120fps GPU-accelerated animations with Motion.dev (Framer Motion successor) for React, Next.js, Svelte, and Astro projects. Use when user requests animation, motion, scroll effects, parallax, hero animations, gestures, drag interactions, spring physics, whileHover effects, whileInView animations, animated UI, micro-interactions, page transitions, or layout animations. Generates production TypeScript/JSX code with accessibility (prefers-reduced-motion) and performance validation (≥60fps). Supports entrance animations, gesture interactions (hover/tap/drag), scroll-based reveals, and layout transitions using spring physics and natural timing. Do NOT use for CSS-only transitions (use native CSS), static sites without JavaScript, Vue animations (use motion-v variant instead), or SVG/Canvas complex animations (GSAP better suited).

NeverSight/learn-skills.dev

Static artifact craft skill for self-contained HTML/CSS/JS documents: docs, sheets, dashboards, explainers, slides, tools, and landing pages. Use when the user asks for a durable, openable, shareable web deliverable they'll keep or hand off — a report, a dashboard, a slide deck, a data table, a page. Local folder first, temporary public link via tunnel (localhost.run), optional durable publish to Surge, GitHub Pages, or Cloudflare. Not for quick look renders, inline snippets, or throwaway scratch. Not for SPA frameworks, backend APIs, database apps, or production product UI.

NeverSight/learn-skills.dev

FFmpeg commands for video/audio conversion, trimming, compression, and processing. Use when user mentions "ffmpeg", "convert video", "compress video", "extract audio", "trim video", "gif from video", "video codec", "transcode", "screen recording", "merge videos", "video to mp4", "reduce file size", or any media processing task.

NeverSight/learn-skills.dev

Vim keybindings, motions, text objects, and operators for efficient text editing. Use when user asks about "vim commands", "vim motions", "text objects", "vim keybindings", "vim cheat sheet", "learn vim", "vim in VS Code", or any Vim editing tasks.

NeverSight/learn-skills.dev

Grok Build ONLY. Turn a 2D character still into smooth animation sprites via image_gen/image_edit base → image_to_video (6s/10s run-in-place) → ffmpeg frames → magenta chroma-key → dense sampled sprites (strip/grid/GIF). Use when the user wants video-to-sprite, motion capture from generated video, smoother run/walk cycles from dense frames, or runs /video2dsprite. Do NOT use on Codex/Claude — only Grok Build has image_to_video. Prefer generate2dsprite for crisp pixel sheets without video.

相关技能