他强行合并了这两个积怨已久、甚至文化互斥的部门,成立了全新的 Google DeepMind。这一刀砍下去,不仅终结了长达十年的内部冷战,彻底终结了“西瓜计划”的所有可能性,更重要的是确立了唯一的权力中心——戴米斯·哈萨比斯(Demis Hassabis)。这位DeepMind的创始人,从原本的“藩王”晋升为整个谷歌AI帝国的最高指挥官。这是一次“集权运动”,哈萨比斯虽然获得了最高指挥权,但也同时失去了寻求独立的法理基础,在未来必须为谷歌的商业产品(下文即将揭晓——虽然现在的你已经知道了)服务。
在Gemini的加持下,谷歌搜索不再仅仅抛给你十个蓝色链接让你自己去翻找。新的 Deep Research 功能就像派遣了一个不知疲倦的研究员。当你询问“2025年日本旅游攻略,避开红眼航班,预算2万”时,Gemini会像一个真正的智能体(Agent)一样,在后台自动拆解出数百个子查询,阅读几十篇游记,比对航班价格,最后直接生成一份带有引用来源的决策报告。
Human-in-the-loop, Crafted by Intelligence. > A professional AI workflow tool designed for long-form, deep content creation. Rejecting the mediocrity of “one-click generation,” it uses a professional pipeline of Material Parsing -> Outline Building -> Sectional Writing to let AI truly become your creative partner.
When generating deep articles over 5,000 words (such as industry analysis, interview transcripts, or long-form features), asking AI to “write an article” directly often leads to logical gaps, hallucinations, or bland styles.
Many AI tools are poorly designed—not targeting anyone specific, but most are garbage—completely failing to fit real workflows. Thus, I was forced to build my own.
Ink & Craft adopts a Human-in-the-loop mode:
Deep Listening: Not just transcribing audio, but understanding context and distinguishing speakers.
Deep Thinking: Utilizing Gemini 3 Pro’s Thinking Mode to conceptualize before writing.
Steady Writing: Ensuring rigorous logic and eliminating hallucinations through “Director’s Annotation” and “Sectional Generation”.
🚀 Quick Start Guide
Step 1: Material Production
(Note: This step is optional. If you already have an idea, thought, outline, or document, you can skip to Step 2 and upload it there.)
Scenario: You have a meeting recording, expert interview, or chaotic voice memos.
Upload File: Supports MP3, WAV, M4A, MP4.
AI Transcription & Diarization:
The system automatically performs Diarization, distinguishing who is speaking.
Streaming Output: Watch text generate in real-time like Matrix code.
Sync Editing:
Click any text segment to jump the audio to that timestamp.
Found an error? Edit the text box directly.
Smart Review:
Click “Start Review”, and AI scans the text, highlighting potential homophone errors or logical contradictions in red.
Export: Generate structured JSON data (with timestamps) for the next steps.
Step 2: Outline Builder
Scenario: You need to define the skeleton and soul of the article.
Reference Material:
System automatically reads the transcript from Step 1 (Optional).
You can also upload PDF, Word, or images as supplements. AI analyzes all your uploaded documents/ideas even without a transcript.
Define Style & Tone:
Professional: Rigorous logic, suitable for research reports.
Interview Note: Structured Q&A, removing filler words, retaining original meaning.
Witty / KOL: Rejecting AI-speak, explaining professional logic like a seasoned blogger using “human language”.
Storytelling: Hero’s Journey structure, suitable for feature stories.
Prompt Tuning: Modify the AI’s System Prompt to inject your unique requirements.
Generate Outline:
AI generates a Markdown outline containing a [Director’s Annotation].
This guide contains metadata like core style and target audience, serving as the “genes” for the next writing step.
Step 3: Article Writing
Scenario: Fleshing out the skeleton into a full-bodied long-form article.
Configuration:
Split Granularity: Choose to split by H1 (coarse) or H2/H3 (fine). Coarse granularity is recommended for long texts to ensure continuity.
Global Instructions: Input extra requirements for the entire article.
Sectional Closed-Loop Generation:
Click the chapter list on the left to generate one by one, or click “Generate Full”.
Anti-Hallucination Mechanism: When generating each chapter, AI looks at the complete material and style guide through a “rearview mirror,” ensuring it stays on topic and doesn’t fabricate facts.
Refinement:
Not satisfied with a section? Click the “Refresh” icon on that chapter.
Tell AI: “This tone is too stiff, add a vivid example,” and it will rewrite that paragraph precisely.
Full Format Export:
Supports Markdown export.
Supports Word (.docx) export, perfectly preserving heading levels and formatting, ready for publishing.
💡 Advanced Tips
1. Director’s Annotation Mode
When generating the outline in Step 2, AI creates metadata at the top of the document like this:
[Director's Annotation]
Core Style: Witty / KOL
Tone Requirements: No "First/Second", use metaphors
Tip: You can manually edit this metadata before starting Step 3! The subsequent AI writing will strictly follow your modified instructions.
2. Deep Thinking (Thinking Mode)
We integrated Google’s latest Gemini 3 Pro model with Thinking Config enabled.
When generating outlines or complex chapters, you will see a “Thinking…” status.
This means AI is performing logical deduction in the background, significantly reducing logical loopholes.
3. Global Debug Console
Click the terminal icon (>_) at the bottom left to see every Prompt sent to AI and the raw AI Response. This is an excellent debugging tool for Prompt Engineering enthusiasts.
FAQ
Q: How long of a recording is supported?
A: Theoretically supports hours of recording. Gemini has a massive Context Window and can “read” the entire interview at once.
Q: Will progress be lost if the page refreshes?
A: No. Progress in all modules is automatically saved in your local browser (LocalStorage).
Q: Why does the generated article sometimes repeat itself?
A: Try adjusting “Split Granularity” to “By H1”. Too fine a granularity causes AI to lack context perception, leading to repetition.