How I Use AI to take perfect notes...without typing

Workflow Overview and Requirements

• An automated workflow can be built to record voice notes, convert them into text, and send the resulting content to a Notion database 0s.
• The automated process generates a full transcript, a summary, a list of main points, and a list of action items from the recorded audio 7s.
• OpenAI's Whisper model is used to transcribe the audio file into text 27s.
• ChatGPT processes the transcript to generate summaries and lists based on specific prompts provided by the user 33s.
• Prompts can be customized to produce different outputs, such as drafting a blog post instead of a standard note 42s.
• This workflow serves as a bridge between personal thoughts and a "second brain" system in Notion, eliminating the need for manual typing on a mobile device 1m15s.
• The "Ultimate Brain" template for Notion includes a dedicated section for voice notes that stores AI-transcribed content, including summaries and transcripts 1m45s.
• Four primary tools are required to build this automation: a Notion account with a notes database, an OpenAI account for access to the Whisper and ChatGPT APIs, a cloud storage provider like Google Drive or Dropbox, and a Pipedream account 2m25s.

Pipedream Automation Setup

• Pipedream serves as the automation builder used to connect various applications and trigger workflows automatically when a new audio file is uploaded to Google Drive 0s.
• The high-level workflow involves recording voice notes on a phone, uploading them to cloud storage like Google Drive or Dropbox, processing them through the Whisper API and OpenAI's ChatGPT API, and finally sending the output to a new page in Notion 12s.
• Behind the scenes, the automation triggers when a file hits a designated "watch folder" in cloud storage, at which point the file is downloaded to temporary storage 38s.
• The technical process includes sending the audio file to Whisper for transcription, sending the resulting text to ChatGPT for summarization, formatting the content, and creating a new page in Notion 48s.
• To begin building, a user must create a Pipedream account and initiate a new workflow, which functions similarly to platforms like Zapier or make.com 1m12s.
• The workflow is named "speech to text to notion" and begins with a trigger step, which cannot be renamed 1m32s.
• The trigger is configured by selecting Google Drive and choosing the "new files" action, which emits an event whenever a file is added to a linked account 1m48s.
• The trigger is restricted to a specific folder, such as an "audio upload test" folder, rather than monitoring the entire Google Drive account 2m15s.
• After setting up the trigger, the system requires a test event, which can be generated by manually uploading an audio file to the specified Google Drive folder 2m35s.

Data Handling and File Processing

• Pipedream displays a green success message when a step completes, providing an exported object that contains properties accessible for use in subsequent automation steps 0s.
• When referencing data from an automation run, it is preferable to use the "copy path" option rather than "copy value" to ensure the automation dynamically references the specific data associated with each new run 15s.
• To process a file with OpenAI's Whisper, the file must first be downloaded from Google Drive into Pipedream's temporary directory because Whisper cannot directly access the user's Google Drive 45s.
• Pipedream uses a specific syntax involving double brackets to reference properties within the "steps" object, allowing users to drill down into nested objects like the trigger event to retrieve specific identifiers 1m25s.
• When configuring the "download file" action in Pipedream, the file ID from the trigger event is used to identify which file to download 1m45s.
• To avoid errors caused by mismatched file formats, the destination file path in the temporary directory should dynamically reference the file extension property from the trigger event rather than using a hardcoded extension like ".mp3" 2m15s.
• Whisper supports various audio file formats, including M4A and MP3, allowing for dynamic file extension settings within the workflow 0s.
• Successful file processing results in a return value that includes the file name and MIME type, confirming the file is stored in temporary storage 12s.

OpenAI Whisper Transcription Configuration

• To transcribe audio, the OpenAI ChatGPT app is added to the workflow, and the "create transcription" action is selected 25s.
• Users without an existing OpenAI account can sign up at platform.openai.com, where new accounts typically receive $5 in free tokens that expire after three months 42s.
• The $5 credit is estimated to cover approximately 12 hours of audio transcription and summarization 55s.
• Account management, including upgrading to a paid plan and setting monthly spending limits, is handled through the billing overview section under the personal menu 1m8s.
• A secret API key must be generated in the user API keys section of the OpenAI dashboard and then pasted into the Pipe Dream workflow 1m23s.
• The audio upload type is set to "file," and the file path is defined by combining the temporary directory path (/tmp/recording) with the file extension retrieved from the trigger step 1m42s.
• Optional fields for the transcription step are generally not required, though they may be utilized in subsequent steps involving ChatGPT 2m15s.
• Testing the transcription step produces a return value containing the full text of the audio file 2m25s.
• Errors indicating that a recording no longer exists may occur during the building process because files in the temporary directory are not stored for an extended period 2m40s.

Workflow Troubleshooting and Optimization

• If a workflow fails, uploading a new file to Google Drive and re-testing the download step can resolve the issue. 0s
• Pipe Dream workflows have a default timeout of 30 seconds, which may be insufficient for processing long files through Whisper. 5s
• To prevent timeouts, users can adjust the execution control settings in the Pipe Dream menu to increase the timeout value, such as setting it to 180 seconds. 15s
• A written tutorial is available that provides additional details regarding Pipe Dream settings, pricing, and an alternative "code-heavy" method for building workflows. 30s

ChatGPT Prompt Engineering

• After obtaining a transcription, an OpenAI ChatGPT step can be added to the workflow to summarize the text or create lists. 55s
• When configuring the ChatGPT API step, users can select the GPT-3.5 Turbo model, noting that specific versions may be frozen as of March 1, 2023. 1m15s
• The quality of ChatGPT output depends on the prompt, which consists of three components: the query (user message), the context (the transcript or data to be analyzed), and the system instructions. 1m35s
• While the website version of ChatGPT combines user messages and system instructions, the API treats them as separate parameters. 2m0s
• System instructions define how the model should respond regarding formatting and are an effective place to provide examples to improve output quality. 2m15s
• Pre-written prompts for the user message field can be sourced from the associated written tutorial. 2m35s
• A specific prompt for summarizing transcripts with ChatGPT is available in a code block within the provided tutorial resources. 0s
• The prompt instructs ChatGPT to generate a title under 15 words and utilizes specific delimiters, such as "--summary--" and "--additional info--", to facilitate the parsing of the output into separate data segments. 22s
• These delimiters allow for the automated placement of the title, summary, and various lists into distinct fields on a Notion page. 42s
• The prompt requests that the summary be formatted as a heading one and includes requirements for lists such as main points, action items, follow-up questions, and potential arguments. 1m5s

AI Output Formatting and Consistency

• The transcript is dynamically linked to the ChatGPT step by referencing the transcription property from the previous step in the workflow. 1m25s
• System instructions are used to ensure consistent output by directing the assistant to communicate exclusively in Markdown formatting. 1m55s
• Providing example formatting within the system instructions serves as an additional measure to ensure the output adheres to the desired structure and includes the required delimiters. 2m15s
• A temperature setting of 0.2 is applied to the ChatGPT configuration to prioritize straightforward, consistent, and well-formatted output over creative generation. 2m35s
• A temperature setting of 0.2 is recommended for the AI configuration, as it provides effective results. 0s
• The AI output is structured as an array containing a content property, which includes the title, summary, additional information, and main points formatted in Markdown. 7s
• Markdown formatting allows for seamless integration into Notion, where the content retains its original structure, such as heading twos and bulleted lists. 23s
• PipeDream is capable of parsing Markdown and transmitting it directly to Notion. 36s

Data Parsing with Node.js

• The current workflow output lacks the transcript and includes the title within the page content rather than the page title, necessitating an additional processing step. 43s
• Adding a Node.js code step to the PipeDream workflow allows for more elegant data handling and reduces the number of required OpenAI API calls, which would otherwise be necessary if processing the title, summary, and points separately. 55s
• PipeDream provides an execution environment that manages security and authentication while allowing users to write custom JavaScript or Python code. 1m15s
• It is recommended to rename the Node step to "formatter" before adding code, as this name is used in the path for referencing values in future steps; changing the name later would break those references. 1m45s
• A pre-written code block is available in a provided guide to facilitate the splitting of the title, summary, transcript, and additional information into distinct properties. 2m5s
• Testing the formatter step results in an object that successfully separates the content into four distinct properties: title, transcript, summary, and additional info. 2m25s
• A specific code block is utilized to format transcripts by breaking them into paragraphs of no more than three sentences each, separated by blank spaces 0s.
• This formatting ensures that Pipedream sends each paragraph to Notion as an individual text block rather than a single, large wall of text 0s.

Notion Integration and Page Creation

• To finalize the process, a final step is added to the workflow by selecting Notion as the application and choosing the "create page from database" action 15s.
• The Notion account must be authenticated, and the specific notes database must be selected to ensure Pipedream has the necessary access 25s.
• If a database does not appear in the Pipedream selection list, access can be granted manually by navigating to the database's three-dot menu, selecting "add connections," and searching for Pipedream 1m5s.
• Once the database is selected as the parent database ID, the page content can be defined using Markdown syntax to dynamically structure the layout 1m35s.
• A code block can be copied and pasted into the page content field to dynamically reference values such as the summary, transcript headings, the transcript itself, and additional information 1m55s.
• Optional fields allow for the configuration of page titles and additional properties, such as setting page icons or covers via the "Meta Types" field 2m15s.
• It is recommended to enable specific property types in the database before configuring the Meta Types 2m35s.
• Configuration of the database properties involves setting the title to the name property and assigning a specific value of "AI transcription" to the type property 0s.
• Users can customize the entry icon by selecting an emoji, with the option to search for specific symbols using keyboard shortcuts like the Windows key and semicolon 15s.
• The title property is configured by dynamically referencing an object, which requires copying the path from the formatter success message and pasting it into the title field 30s.
• After selecting the "AI transcription" option from the Notion database, the workflow can be tested to ensure the transcript and summary appear in the designated filtered view 48s.
• A successful test results in the creation of a new page in Notion containing the transcript, summary, main points, action items, and follow-up questions 1m5s.

Advanced Workflow Deployment and Scaling

• Deploying the workflow enables it to automatically trigger whenever a new audio file is uploaded to the connected Google Drive folder 1m22s.
• The no-code version of this workflow is subject to a 25-megabyte file size limit imposed by Whisper 1m33s.
• A code-heavy version of the tutorial is available for users who need to transcribe longer audio files, such as hour-long recordings or podcast episodes 1m42s.
• The code-heavy version requires users to copy and paste code into Pipedream, with the option to use ChatGPT for explanations or modifications 1m52s.

Notion Resources and Community Support

• The Ultimate Brain template for Notion provides a pre-built system for note-taking, task management, project management, goal tracking, and recipe tracking 2m15s.
• Ultimate Brain includes a dedicated notes dashboard designed for mobile use and provides access to a support team and a community of users for assistance and customization 2m35s.
• A template is available for acquisition at thomasjfrank.com/brain, with a discount code provided in the video description 0s.
• Interested individuals can subscribe to the Notion Tips newsletter to receive notifications regarding new Notion tutorials and template releases 7s.
• A free Notion Fundamentals course is accessible on the website, covering topics ranging from page basics and writing systems to databases 14s.
• New lessons for the Notion Fundamentals course are currently under development 24s.
• Viewers can sign up for the Notion Tips email newsletter directly through the Notion Fundamentals page 27s.
• Questions regarding the content can be posted in the comment section or directed to the Twitter account @TomFrankly 31s.
• The workflow discussed is intended to assist individuals who frequently generate ideas while on the go 45s.