Turn trip receipts into an expense report and triage files
I haven’t opened ChatGPT in the last 30 days because of Claude Cowork:. Task: I haven’t opened ChatGPT in the last 30 days because of Claude Cowork: I haven’t opened ChatGPT for almost a month. First, I’ll cover 4 things that made me switch to the new Claude Cowork. First, I’ll cover 4 things that made me switch to the new Claude Cowork. Then I’ll show you 5 things I’ve built with it in the last 3 weeks. Then I’ll show you 5 things I’ve built with it in the last 3 weeks. Anyone can replicate what I did. I don’t know how to code. I am non-technical. Give me 7 minutes. Follow my lead: #1. It actually follows instructions Here’s the problem with ChatGPT. You give it a detailed prompt. 500 words of context. Specific requirements. Formatting rules. Things you want. Things you don’t want. It ignores half of it. You remind it. It apologizes. Then ignores the other half. I got tired of fighting. With Claude Cowork, I upload my voice profile. Over 2,000 words describing how I write, what I hate, phrases I never use, structures I prefer. I upload my copywriting rules. I upload reference documents. It reads all of it. It follows all of it. When I write with Claude Cowork, the output sounds like me. Not “AI trying to sound like me.” Me. This matters more than you think. The quality of your output depends on how much context you can give. If the AI ignores your context, you’re stuck with generic results. Forever. Play this video to understand how it works: Install Claude (the app) here. Install Claude (the app) here. You need a Pro account (seventeen dollars a month). You need a Pro account (seventeen dollars a month). When you open Claude, simply go to the tab “Cowork”. When you open Claude, simply go to the tab “Cowork”. #2. I’m finally out of the infinite loop. You know the feeling. You’re 45 minutes into a complex task. ChatGPT starts repeating itself. Or contradicting what it said earlier. Or getting stuck in a loop where it keeps apologizing and producing the same wrong output. You start over a new chat. Or work around it. Or give up. Now I’ve had 2-hour sessions in Claude Cowork. Complex back-and-forth. Revisions. Pivots. New directions mid-conversation. 1- It doesn’t lose the thread. 2- It doesn’t get trapped. 3- It remembers what we decided 90 minutes ago and builds on it. 4- It creates a form to ask me about a potential new direction (since I’m not satisfied with the output). Real work takes time & multiple pivots. If your AI breaks after 20-minute sessions, you can only do small tasks. That’s a ceiling on what’s possible. My favorite trick (when eveything is going wrong) is this prompt: You understood that right. Claude Cowork is the one prompting me. Play this video to see it in action: #3. No memory is the best somehow. Yes, this is a feature. Not a bug. ChatGPT remembers your past conversations. Sounds useful. But it creates what I call overfitting. It tries too hard to match what you said last week. It makes assumptions based on old context that no longer applies. It gets weird in ways that are hard to diagnose. And I’m not the only one hating it: But Claude Cowork starts fresh every time. No baggage. No assumptions from three weeks ago. If I want Claude to know something, I give it that context explicitly. In this conversation. For this task. With my favorite markdown files to support it. I decide what matters. Not some algorithm guessing based on a conversation I don’t even remember having. Sure, it would have been awesome to have an AI knowing exactly when to pull the right information from past conversations. But it simply does not work like this. So instead of fighting it, or not using it because it can’t, simply use AI the way its best. And this is when you control the full context before you start the task. You’re sick to introduce yourself? Create a markdown file with everything about you, and upload it to any AI at the start of every conversation (when necessary). So far Claude Cowork can: Follow (long) instructions properly. Follow (long) instructions properly. Prompt itself (so I don’t have to). Prompt itself (so I don’t have to). Avoid overfitting (with memories). Avoid overfitting (with memories). Time for #4: Claude create files directly. #4. It creates actual files, not just text. This is the big one. ChatGPT gives you text. You copy that text. You paste it into another tool. You format it yourself. Every. Single. Time. Claude Cowork creates the actual files. I ask for a presentation. I get slides. Real slides.; tools: Claude Cowork. Human role: The original post does not say how a person stayed involved.; limitations: Requires a Claude Pro account ($17/month) and the Cowork tab. Image quality affects extraction accuracy; blurry receipts require manual intervention. The source is a promotional blog post, so performance claims may not reflect enterprise-scale reliability or security compliance..
Tools
Published on Agent Work Atlas:
Overview
I haven’t opened ChatGPT in the last 30 days because of Claude Cowork:
Workflow
I haven’t opened ChatGPT in the last 30 days because of Claude Cowork: I haven’t opened ChatGPT for almost a month. First, I’ll cover 4 things that made me switch to the new Claude Cowork. First, I’ll cover 4 things that made me switch to the new Claude Cowork. Then I’ll show you 5 things I’ve built with it in the last 3 weeks. Then I’ll show you 5 things I’ve built with it in the last 3 weeks. Anyone can replicate what I did. I don’t know how to code. I am non-technical. Give me 7 minutes. Follow my lead: #1. It actually follows instructions Here’s the problem with ChatGPT. You give it a detailed prompt. 500 words of context. Specific requirements. Formatting rules. Things you want. Things you don’t want. It ignores half of it. You remind it. It apologizes. Then ignores the other half. I got tired of fighting. With Claude Cowork, I upload my voice profile. Over 2,000 words describing how I write, what I hate, phrases I never use, structures I prefer. I upload my copywriting rules. I upload reference documents. It reads all of it. It follows all of it. When I write with Claude Cowork, the output sounds like me. Not “AI trying to sound like me.” Me. This matters more than you think. The quality of your output depends on how much context you can give. If the AI ignores your context, you’re stuck with generic results. Forever. Play this video to understand how it works: Install Claude (the app) here. Install Claude (the app) here. You need a Pro account (seventeen dollars a month). You need a Pro account (seventeen dollars a month). When you open Claude, simply go to the tab “Cowork”. When you open Claude, simply go to the tab “Cowork”. #2. I’m finally out of the infinite loop. You know the feeling. You’re 45 minutes into a complex task. ChatGPT starts repeating itself. Or contradicting what it said earlier. Or getting stuck in a loop where it keeps apologizing and producing the same wrong output. You start over a new chat. Or work around it. Or give up. Now I’ve had 2-hour sessions in Claude Cowork. Complex back-and-forth. Revisions. Pivots. New directions mid-conversation. 1- It doesn’t lose the thread. 2- It doesn’t get trapped. 3- It remembers what we decided 90 minutes ago and builds on it. 4- It creates a form to ask me about a potential new direction (since I’m not satisfied with the output). Real work takes time & multiple pivots. If your AI breaks after 20-minute sessions, you can only do small tasks. That’s a ceiling on what’s possible. My favorite trick (when eveything is going wrong) is this prompt: You understood that right. Claude Cowork is the one prompting me. Play this video to see it in action: #3. No memory is the best somehow. Yes, this is a feature. Not a bug. ChatGPT remembers your past conversations. Sounds useful. But it creates what I call overfitting. It tries too hard to match what you said last week. It makes assumptions based on old context that no longer applies. It gets weird in ways that are hard to diagnose. And I’m not the only one hating it: But Claude Cowork starts fresh every time. No baggage. No assumptions from three weeks ago. If I want Claude to know something, I give it that context explicitly. In this conversation. For this task. With my favorite markdown files to support it. I decide what matters. Not some algorithm guessing based on a conversation I don’t even remember having. Sure, it would have been awesome to have an AI knowing exactly when to pull the right information from past conversations. But it simply does not work like this. So instead of fighting it, or not using it because it can’t, simply use AI the way its best. And this is when you control the full context before you start the task. You’re sick to introduce yourself? Create a markdown file with everything about you, and upload it to any AI at the start of every conversation (when necessary). So far Claude Cowork can: Follow (long) instructions properly. Follow (long) instructions properly. Prompt itself (so I don’t have to). Prompt itself (so I don’t have to). Avoid overfitting (with memories). Avoid overfitting (with memories). Time for #4: Claude create files directly. #4. It creates actual files, not just text. This is the big one. ChatGPT gives you text. You copy that text. You paste it into another tool. You format it yourself. Every. Single. Time. Claude Cowork creates the actual files. I ask for a presentation. I get slides. Real slides.
Reported result
The workflow reduced the task time from an estimated 2 hours to 10 minutes. The AI successfully logged 44 receipts automatically but flagged 3 blurry ones for manual review. No broader business impact or error rates beyond these specific instances are reported.
Human role
The original post does not say how a person stayed involved.
Limitations
Requires a Claude Pro account ($17/month) and the Cowork tab. Image quality affects extraction accuracy; blurry receipts require manual intervention. The source is a promotional blog post, so performance claims may not reflect enterprise-scale reliability or security compliance.
Try it
Act as an expense processing agent. I will provide a folder path [replaceable input] containing trip receipt images/PDFs. First, perform a dry run on 3 sample receipts to verify extraction accuracy. If successful, process the remaining files. For each receipt, extract Date, Vendor, Amount, and Category. If an image is unclear, flag it in the output rather than guessing. Generate a single Excel file with columns: Date, Vendor, Amount, Category, Status (Processed/Flagged). Include a summary sheet with total expenses by category. Do not submit or email the report; save it to [output location] and ask for my confirmation before finalizing. If any receipt is missing critical data, list it in a 'Needs Attention' tab.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.