# Turn a voice memo, a photo and a spreadsheet into a proposal draft in 90 minutes Source: https://getstartedwithai.net/tutorials/multi-modal-workflow Site: Get Started With AI Tagline: Build AI automations that actually run your business. Audience: solo professionals, freelancers, consultants and AI implementers in training Contact: hello@getstartedwithai.co.uk This Markdown companion is generated from the prerendered page for AI crawlers. # Turn a voice memo, a photo and a spreadsheet into a proposal draft in 90 minutes Most professionals have three types of raw material sitting on their phone at the end of a client site visit: a voice note with the verbal brief, a photo of the whiteboard or the broken machine, and a CSV or exported spreadsheet of the relevant numbers. The standard workflow is to manually transcribe all three, stitch the pieces together, then write the proposal from scratch. A multimodal AI workflow replaces the transcription and stitching work entirely. A modern vision-capable model such as Claude 3.5 Sonnet, ChatGPT-4o or Gemini 1.5 Pro can read text, interpret an image and process a data file inside a single conversation. The gap most practitioners miss is that you do not need a separate transcription tool, a separate OCR tool and a separate data tool. One model handles all three inputs together. This tutorial walks through the complete workflow. You end up with a structured proposal draft, a reusable prompt template, and a service you can package for clients at between 400 and 1,200 pounds per engagement. ## Prep before you build (5 minutes) - A vision-capable AI account. Claude 3.5 Sonnet (claude.ai), ChatGPT-4o (chat.openai.com) or Gemini 1.5 Pro (gemini.google.com) all work. The free tiers support file uploads for testing. A paid tier unlocks longer uploads and higher rate limits. - A voice memo. Any recording from your phone voice app works. iOS Voice Memos, Android Recorder, WhatsApp audio all export as audio files. - An image file. A whiteboard photo, a photo of a printed document, a screenshot of a Figma board, a product photo. JPEG or PNG both work. - A CSV or spreadsheet. An exported row from a CRM, a client-supplied data file, a quick export from Xero or QuickBooks, or a manually built table pasted as plain text. The model reads CSV natively. For Excel files, save as CSV first. - A proposal template. One page is enough. You need a Situation section, an Approach section, a Deliverables section and an Investment section. You can paste this into the prompt. ## Step 1: Transcribe the voice memo (10 minutes) Open your AI account. Upload the audio file directly to a new conversation. 1. Prompt: Transcribe this audio file. Return the full transcript without summarising or paraphrasing. Flag any words you are uncertain about with [unclear] so I can check them. 2. Read the transcript back against the original. Fix any flagged words. 3. Save the clean transcript to a scratch document. This becomes the verbal brief input for the assembly step. ## Step 2: Extract the image content (10 minutes) Start a new conversation or continue the existing one. Upload the photo or image. 1. Prompt: Describe everything you can read or see in this image. If it contains text, return it verbatim. If it contains a diagram or whiteboard, describe the structure and content in plain text. Do not interpret or summarise. Describe. 2. Review the output. For whiteboard photos with poor lighting, ask the model to flag anything unclear. You will catch about 85 to 90 percent of the information in a standard whiteboard photo. 3. Supplement the output with your own notes for anything the model flags as unclear. ## Step 3: Parse the spreadsheet data (10 minutes) Upload the CSV file or paste the table as plain text. 1. Prompt: Read this data. Return the column headers and what each column appears to measure, a three-sentence plain-English summary of what the data shows, and the three numbers or trends most likely to matter in a proposal context. Do not add interpretation beyond what the data supports. 2. Review the output. Numbers from the CSV now land in a readable summary you can paste directly into the Situation and Investment sections. 3. This step replaces manual data lookup for the proposal. The factual backbone is done. ## Step 4: Combine all three into a proposal draft (20 minutes) In the same conversation, or a fresh one with all three prior outputs pasted in, run the assembly prompt. 1. Prompt: Here are three inputs for a proposal draft: the voice memo transcript, the image content description, and the spreadsheet summary. My proposal template has four sections: Situation, Approach, Deliverables, Investment. Draft the proposal using only facts from the three inputs above. Where data is missing for a section, mark it [FILL IN] rather than inventing. Use plain English, not consultant jargon. Maximum 600 words. 2. The model will cross-reference the three inputs and populate the template. Sections built from the data CSV will be specific. Sections that rely only on the voice memo may need more detail. 3. Any [FILL IN] markers tell you exactly what to ask the client before sending. ## Step 5: Edit and strengthen the draft (20 minutes) Read the draft out loud. Cut anything that does not earn its place. The voice memo transcript often contains filler that the model reproduces faithfully. 1. Run a second prompt to tighten the Situation section: Rewrite the Situation section below to be 30 percent shorter without losing any facts. Cut hedging language throughout and ensure every sentence states something specific rather than something approximate. 2. Fill in every [FILL IN] marker. Either pull the missing information from memory, or note it as a follow-up question for the client. 3. Do a final read for jargon. Replace any phrase you would not say out loud on a client call. ## Step 6: Save the reusable prompt template (5 minutes) 1. Save the assembly prompt from Step 4 to your Claude Project instructions or ChatGPT Saved Prompts. 2. The next time you run this workflow, you paste in the three transcript blocks and the model does the assembly in one step. 3. Add a note of the client, date and project name to each saved transcript before archiving. Three months from now you will want to find them. ## What you ship A structured four-section proposal draft built entirely from the raw inputs, with no manual transcription or data re-entry. A saved assembly prompt you can reuse for every site-visit engagement. A reusable workflow you can offer as a packaged service to clients who do proposal-heavy work. Consultants running this workflow report cutting proposal turnaround from four to six hours to 90 minutes or less, including the editing pass. ## What to build next Add a Whisper API transcription step to handle longer recordings or multiple speakers. Pair this with the chain prompts tutorial to split the assembly across stages where the brief is complex. For clients with recurring proposal work, wrap the workflow in a Make or Zapier scenario that triggers on a new audio file upload. ## Related builds More tutorials in Automation. Each one runs in 60 to 90 minutes and ships a finished asset. ### Capture every lead automatically: a 30-minute Zapier build 30 to 60 minutes. Solo Pro. [Open tutorial](/tutorials/first-zapier-automation-30-min) ### Decide between native AI Steps and a Custom GPT in 10 minutes 60 to 90 minutes. Solo Pro. [Open tutorial](/tutorials/zapier-ai-features-vs-custom-gpt) ### Run automations Zapier cannot: the first Make scenario 60 to 90 minutes. Solo Pro. [Open tutorial](/tutorials/first-make-scenario-visual) ### Self-host the automation stack: when n8n is worth the effort 60 to 90 minutes. Solo Pro / Implementer. [Open tutorial](/tutorials/n8n-self-hosted-when-worth-it) ## Questions people ask before they start ### How accurate is the voice transcription? For clear audio at normal speaking pace, modern models hit 95 to 97 percent accuracy on standard English. Accuracy drops with strong regional accents, background noise, or technical jargon. Read the transcript back once and fix flagged words before passing it to the assembly prompt. ### Can I use a photo taken on a phone rather than a proper scan? Yes. A well-lit whiteboard photo taken from directly in front of the board at 1 to 2 metres reads well. Avoid angle shots where the writing runs to a corner. If the photo is borderline, run it through the image extraction step first and check the output before combining. ### What if my spreadsheet has 50 columns? Narrow it first. Delete columns that are not relevant to the proposal before uploading. The model handles large tables, but the summary prompt works best when the data is scoped to the proposal topic. Ten relevant columns beat 50 mixed ones. ### How much should I charge to set this up for a client? For a single workflow setup with a saved prompt template and a 30-minute handover session, charge between 400 and 700 pounds. If the client wants the workflow wrapped in an automation trigger, that is a separate build at 800 to 1,500 pounds. Implementers running this as a recurring service for proposal-heavy businesses charge 250 to 400 pounds per month as a managed workflow fee. ### Do I need a paid AI account? For testing and occasional use, the free tier works. Claude.ai free allows file uploads but limits conversation length. ChatGPT free allows image uploads but not audio. For a production workflow, a paid tier at around 16 to 20 pounds per month removes the rate limits and unlocks full file-type support across audio, image and data.