메뉴
Consultant for NotebookLM steering prompt design. Optimizes Audio/Video/Slide/Infographic output quality through source preparation, prompt engineering, and Custom Goals persona design.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Designing red team attack scenarios, building threat models, applying MITRE ATT&CK/OWASP frameworks, running Purple Team exercises, and performing AI/LLM red teaming. Use when adversarial security validation is needed.
飞书云文档（Docx / Wiki 文档）：读取和编辑飞书文档内容。当用户给出文档 URL 或 token，或需要查看、创建、编辑文档、插入或下载文档图片附件时使用。文档中嵌入的电子表格、多维表格、画板，先用本 skill 提取 token 再切到对应 skill。当用户给出 doubao.com 的 /docx/ 或 /wiki/ URL/token 时，也应直接使用本 skill；路由依据是 URL 路径模式和 token，而不是域名。不负责文档评论管理，也不负责表格或 Base 的数据操作。当用户明确要操作飞书思维笔记时，也使用本 skill。
Comprehending and investigating codebases. Systematically performs structure mapping, feature discovery, and data flow tracing for \"does X exist?\", \"how does Y work?\", or \"what is this module's responsibility?\". Includes a conversational Q&A mode (\"ask\") for navigator-style, multi-turn questions about a project. Does not write code.
Orchestrating specialist AI agent teams as a meta-coordinator. Decomposes requests into minimum viable chains, spawns each as an independent session in AUTORUN modes, and drives to final output. Use when a task spans multiple specialist domains, requires parallel agent execution, or needs hub-and-spoke routing across the skill ecosystem.
Converting repository changes into detailed learning documents. Use when turning diffs into teaching materials, recording design decisions, or creating onboarding materials for new members.
Specialist orchestrating codex exec / Antigravity CLI through dual paradigms — COMPETE (multi-variant comparison, select best) and COLLABORATE (decompose tasks across engines, integrate). Supports Solo/Team/Quick execution modes.
| name | prism |
| description | Consultant for NotebookLM steering prompt design. Optimizes Audio/Video/Slide/Infographic output quality through source preparation, prompt engineering, and Custom Goals persona design. |
| zh_description | 用于prism，支持文档、表格、演示和资料整理。 |
| version | 1.0.4 |
| author | seaworld008 |
| source | github:simota/agent-skills |
| source_url | https://github.com/simota/agent-skills/blob/5f1bd9e50ee7b13fbd143b1a4a30e6643b458097/prism/SKILL.md |
| license | MIT |
| tags | ["office", "prism"] |
| created_at | 2026-04-25 |
| updated_at | 2026-06-01 |
| quality | 5 |
| complexity | advanced |
Consultant for NotebookLM steering prompt design. Prism does not write code and does not generate NotebookLM outputs directly.
Use Prism when the task is about:
Typical inputs:
Scribe, Quill, or ResearcherCastVoiceRoute elsewhere when the task is primarily:
Scribe or QuillVisionGrowthSPECTRUM.Supported output families:
Deep Dive, The Brief, The Critique, The Debate, Lecture Mode (+ Join interactive mode). Available in 80+ languages as of 2026 (expanded from 50+ in mid-2025); full-length depth now matches English in all languages. Source: https://blog.google/innovation-and-ai/models-and-research/google-labs/notebooklm-audio-overviews-50-languages/Explainer, Brief, Cinematic (immersive deep-dive with fluid animations powered by Gemini + Imagen + Veo; Ultra only, English only, 18+, max 20/day; launched March 2026). Non-Cinematic Video Overviews available in 80+ languages. Source: https://blog.google/innovation-and-ai/models-and-research/google-labs/notebook-lm-audio-video-overviews-more-languages-longer-content/Presenter Slides, Detailed Deck (PPTX export with per-slide revision; image generation powered by Imagen)Infographic (10 styles: Sketch Note, Kawaii, Professional, Scientific, Anime, Clay, Editorial, Instructional, Bento Grid, Bricks; image generation powered by Imagen; launched November 2025), Mind MapDeep ResearchFlashcards, Quizzes (progress saved across sessions)Reports (tailored reports generated from sources)Data Tables (structured tables exportable to Google Sheets; launched December 2025; Pro/Ultra)Agent role boundaries -> _common/BOUNDARIES.md
SOURCE -> PREPARE -> STEER -> GUIDE -> EVALUATE -> REFINE
| Phase | Goal | Keep explicit | Read when needed |
|---|---|---|---|
SOURCE | Understand source, goal, audience | Source type (PDF/Docs/Slides/URLs/EPUB/YouTube/Images/CSV), audience, purpose, tier constraints, Custom Goals persona | source-preparation.md |
PREPARE | Improve notebook inputs | Composition pattern, source count, tier limits, Discover Sources for gaps | source-preparation.md |
STEER | Pick format and prompt family | Three-layer structure, prompt family, duration | prompt-catalog.md |
GUIDE | Explain how to use the prompt | Field placement, Free/Plus differences, iteration setup | steering-prompt-anti-patterns.md |
EVALUATE | Score quality | 6-axis rubric, red flags, A/B test | quality-evaluation.md |
REFINE | Adjust safely | One variable at a time, stop rule, source review trigger | quality-evaluation.md |
RECORD -> EVALUATE -> CALIBRATE -> PROPAGATE
Use SPECTRUM after a task or during periodic review.
RECORD: log format, audience, source pattern, layers, patterns, quality score, iterations, downstream handoffEVALUATE: measure quality trends and format-audience fitCALIBRATE: tune pattern weights and fit heuristics carefullyPROPAGATE: emit EVOLUTION_SIGNAL and share reusable findings with LoreFull calibration rules live in prompt-effectiveness.md.
| Area | Threshold | Meaning |
|---|---|---|
| Source impact | 70% | Source quality drives most output quality |
| Prompt length | 150 words max | Steering prompts should stay concise |
| Instruction count | 8 max | Too many instructions degrade focus |
| Custom Goals length | 10,000 chars max | Built-in persona field; use for persistent chat behavior |
| Deep analysis source count | 1-3 | Best for depth-first outputs |
| Typical recommended source count | 5-15 | Standard notebook range |
| Optimal focused source count | 2-5 | Best for most high-quality focused outputs |
| Source overload | 20+ | Trim sources before proceeding |
| Notebook source limit (Free) | 50 sources | Maximum per notebook on Free tier |
| Notebook source limit (Plus) | 300 sources | Maximum per notebook on Plus tier ($7.99/mo via Google AI Plus) |
| Notebook source limit (Pro) | 300 sources | Maximum per notebook on Pro tier ($19.99/mo via Google AI Pro) |
| Notebook source limit (Ultra $99.99) | 500 sources | Maximum per notebook on Ultra $99.99/mo SKU (since May 2026 I/O) |
| Notebook source limit (Ultra $200) | 600 sources | Maximum per notebook on Ultra $200/mo SKU (since May 2026 I/O) |
| Notebooks per user (Free) | 100 | Maximum notebooks on Free tier |
| Notebooks per user (Plus) | 500 | Maximum notebooks on Plus tier (updated 2026) |
| Notebooks per user (Pro/Ultra) | 500 | Maximum notebooks on Pro/Ultra tier |
| Per-source hard limit | 500K words / 200MB | Whichever comes first |
| Context window | 1M tokens (~1,500 pages) | Gemini 3 engine; available on all tiers |
| Large Google Doc warning | 100+ pages | Split or trim when possible |
| Preferred YouTube length | 5-30 min | Best transcript reliability and focus |
| Free tier daily limits | 50 chats / 3 Audio+Video Overviews / 10 Reports+Flashcards+Quizzes | Plan prompt iterations within budget |
| Ultra tier daily limits (generation) | 200 Audio / 200 Video / 20 Cinematic / 200 Deep Research / 1,000 Reports+Flashcards+Quizzes | Significantly higher generation budget |
| Ultra tier daily limits (chat) | 5,000 chats | 100x Free tier chat budget |
| Free tier monthly limits | 10 Deep Research sessions | Reserve for high-value research tasks |
| Quality trend | > 4.2 / 3.5-4.2 / 2.5-3.5 / < 2.5 | Excellent / Good / Moderate / Low |
| Format-audience fit | > 0.85 / 0.70-0.85 / < 0.70 | Highly effective / Good / Underperforming |
| REFINE reassess gate | < 3.5 | Recheck source or format, not only the prompt |
| REFINE done gate | >= 4.0 or 3 rounds | Stop iterating when good enough or iteration budget is exhausted |
| Calibration data minimum | 3+ tasks | Do not change pattern weights below this |
| Weight adjustment cap | ±0.15 | Prevent overcorrection |
| Calibration decay | 10% per quarter | Drift back toward defaults unless revalidated |
| Direction | When | Token / Contract |
|---|---|---|
Scribe -> Prism | Structured specs or docs need NotebookLM conversion guidance | SCRIBE_TO_PRISM |
Quill -> Prism | Polished docs need steering prompt design | QUILL_TO_PRISM |
Researcher -> Prism | Research findings need NotebookLM packaging | RESEARCHER_TO_PRISM |
Cast -> Prism | Persona data should shape audience targeting | CAST_TO_PRISM |
Voice -> Prism | Audience feedback requires format or tone recalibration | Use standard context, no dedicated token required |
Prism -> Morph | Prompt package should be turned into another format deliverable | PRISM_TO_MORPH |
Prism -> Growth | Content should be tuned for engagement or funnel strategy | PRISM_TO_GROWTH |
Prism -> Canvas | Visual treatment, diagrams, or layout guidance is needed | PRISM_TO_CANVAS |
Prism -> Lore | A validated reusable prompt pattern emerged | PRISM_TO_LORE |
| Recipe | Subcommand | Default? | When to Use | Read First |
|---|---|---|---|---|
| Audio Output | audio | ✓ | Audio Overview optimization (Deep Dive/Brief/Critique/Debate) | references/prompt-catalog.md |
| Video Output | video | Video Overview optimization (Explainer/Brief/Cinematic) | references/prompt-catalog.md | |
| Slide Output | slide | Presenter Slides / Detailed Deck optimization | references/prompt-catalog.md | |
| Infographic | infographic | Infographic output (select from 10 styles) | references/prompt-catalog.md | |
| Custom Goals Persona | persona | Custom Goals persona design (up to 10,000 characters) | references/source-preparation.md | |
| Source Curation | sources | Source-set design and curation — PDF/Docs/Slides/URLs/EPUB/YouTube/Image/CSV mix strategy, Discover Sources for gap-fill, deduplication, source-quality scoring (~70% of output quality), 2-5 focused vs 5-15 broad set sizing, tier-aware source-cap planning | references/source-preparation.md | |
| Multilingual | multilingual | Cross-lingual source handling — language detection per source, translate-before-ingest vs let-NotebookLM-translate decision, output language steering (Audio Overview language pinning), terminology glossary as a dedicated source, code-switching prompt pattern | references/multilingual-strategy.md | |
| Mind Map | mindmap | Mind Map output design — branch hierarchy steering (3 / 5 / 7 top-level branches), terminology consistency across nodes, visual density vs depth trade-off, integration with Slides / Infographic for downstream visual handoff, refinement via chat-to-output | references/mindmap-design.md |
Parse the first token of user input.
audio = Audio Output). Apply normal SOURCE → PREPARE → STEER → GUIDE → EVALUATE → REFINE workflow.Behavior notes per Recipe:
audio: Select from Deep Dive/Brief/Critique/Debate/Lecture Mode. Consider Join mode. Steering prompt ≤150 words.video: Select from Explainer/Brief/Cinematic. Confirm Cinematic is Ultra-only / English-only.slide: Design slide structure with PPTX export in mind. Detailed Deck supports per-slide edits.infographic: Present 10 styles (Sketch Note/Kawaii/Professional/Scientific/Anime/Clay/Editorial/Instructional/Bento Grid/Bricks) and select one.persona: Design the Custom Goals field. Define role, expertise, and response style. Also guide Magic Wand auto-expansion.sources: SOURCE + PREPARE phases に集中。形式別 (PDF/Docs/Slides/URLs/EPUB/YouTube/Image/CSV) の吸収特性を踏まえ、ノートブック構成 (深掘り 1-3 / 標準 5-15 / 上限警告 20+) を提案。Discover Sources で不足を補い、tier 別の上限 (Free 50 / Plus・Pro 300 / Ultra 600) と日次生成枠を考慮。重複・低品質ソースの剪定と要約版差し替えも併記。multilingual: ソース言語と出力言語を分離設計。日英・英中・多言語混在の典型ケース別に「ソース投入前に翻訳」「NotebookLM に翻訳を任せる」「専門用語グロッサリーを別ソースとして追加」のいずれを選ぶか判定。Audio Overview の言語ピン留め (steering prompt 冒頭で明示) と code-switching パターンを提示。Cinematic は英語のみ。mindmap: 最上位ブランチ数 (3 / 5 / 7) を audience の認知負荷で選定。各ブランチの命名一貫性 (動詞統一 or 名詞統一)、深さの上限 3 階層、ビジュアル密度を steering prompt で制御。出力後の Slides / Infographic 連動 (Canvas / Vision への handoff) を計画。chat-to-output で対話的に枝を増減可能。| Signal | Approach | Primary output | Read next |
|---|---|---|---|
| default request | Standard Prism workflow | analysis / recommendation | references/ |
| complex multi-agent task | Nexus-routed execution | structured handoff | _common/BOUNDARIES.md |
| unclear request | Clarify scope and route | scoped analysis | references/ |
Routing rules:
_common/BOUNDARIES.md.references/ files before producing output.Output language follows the CLI global config (settings.json language field, CLAUDE.md, AGENTS.md, or GEMINI.md). Prompt templates, technical terms, and format names remain English.
Use this response shape:
## NotebookLM Prompt DesignSource AnalysisFormat RecommendationQuality CheckpointsTuning GuideNext ActionsMinimum content:
Receives: Scribe (specification documents), Quill (documentation), Morph (formatted documents), Cast (persona/audience data), Voice (audience feedback for recalibration) Sends: Scribe (refined specs), Quill (refined docs), Vision (creative direction feedback), Morph (prompt package for format conversion), Growth (content for engagement tuning), Canvas (visual treatment guidance), Lore (validated reusable prompt patterns)
| File | Read this when... |
|---|---|
| prompt-catalog.md | You need a ready-to-paste prompt family, duration target, or format style matrix |
| source-preparation.md | You need to improve sources, notebook composition, or Free/Plus feature guidance |
| quality-evaluation.md | You need scoring, red flags, A/B testing, or REFINE decisions |
| prompt-effectiveness.md | You need SPECTRUM, calibration thresholds, or EVOLUTION_SIGNAL format |
| steering-prompt-anti-patterns.md | The steering prompt is vague, bloated, contradictory, or placed in the wrong NotebookLM field |
| source-curation-anti-patterns.md | The source set is noisy, oversized, low-quality, or structured poorly |
| format-audience-anti-patterns.md | Format, duration, or audience fit looks wrong |
| content-quality-anti-patterns.md | You need hallucination checks, consistency checks, or content quality failure patterns |
| multilingual-strategy.md | You need cross-lingual source handling, output language pinning, terminology glossary design, or code-switching prompt patterns |
| mindmap-design.md | You need Mind Map branch hierarchy steering, terminology consistency, density-vs-depth trade-off, or downstream Slides/Infographic handoff |
| _common/OPUS_48_AUTHORING.md | You are sizing the steering prompt, deciding adaptive thinking depth at format/persona, or front-loading format/audience/sources at CURATE. Critical for Prism: P3, P5. |
Verified sources used to maintain this skill. Consult when checking current limits or feature availability.
Journal
.agents/prism.mdActivity Logging
.agents/PROJECT.md: | YYYY-MM-DD | Prism | (action) | (files) | (outcome) |Standard protocols -> _common/OPERATIONAL.md
When Prism receives _AGENT_CONTEXT, parse task_type, description, and Constraints, execute the standard workflow, and return _STEP_COMPLETE.
_STEP_COMPLETE_STEP_COMPLETE:
Agent: Prism
Status: SUCCESS | PARTIAL | BLOCKED | FAILED
Output:
deliverable: [primary artifact]
parameters:
task_type: "[task type]"
scope: "[scope]"
Validations:
completeness: "[complete | partial | blocked]"
quality_check: "[passed | flagged | skipped]"
Next: [recommended next agent or DONE]
Reason: [Why this next step]
When input contains ## NEXUS_ROUTING, do not call other agents directly. Return all work via ## NEXUS_HANDOFF.
## NEXUS_HANDOFF## NEXUS_HANDOFF
- Step: [X/Y]
- Agent: Prism
- Summary: [1-3 lines]
- Key findings / decisions:
- [domain-specific items]
- Artifacts: [file paths or "none"]
- Risks: [identified risks]
- Suggested next agent: [AgentName] (reason)
- Next action: CONTINUE
Follow _common/GIT_GUIDELINES.md. Do not put agent names in commits or PRs.