> ## Content Index
> Fetch the complete content index at: https://www.mindovermoney.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Volume 54: The Instructions That Travel With You Between Tools
- URL: https://www.mindovermoney.ai/how-to-compare-two-ai-tools/
- Published: 2026-10-20T12:00:00.000Z
- Updated: 2026-10-20T12:00:00.000Z
- Description: To find out which AI tool is better for a job you repeat, test both on a small set of cases built from mistakes you have already caught, with the right answer written first. Run each case more than once, and for work others act on, keep the tool that passes every time.
- Author: Santosh Savel
- Tags: Newsletter

While one AI tool helped me build the brief for this issue’s Founder’s Corner, another was running the AI Education workflow, and both drew on the same working methods and context. That setup took a series of small experiments with how I organize my work. Each one built on what I had already learned.

🧭 Founder’s Corner: What you teach AI about your work should carry across tools, and that foundation can start with a project you already know.

🧠 AI Education: Learn how a small set of tests drawn from your own work, run more than once, shows which AI tool handles a task reliably.

✅ 10-Minute Win: Three test cases from mistakes you have already caught, two runs per tool, and a clear call on which assistant keeps the job.

Let’s jump in.

| ![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/eaol-show-square-320.jpg) | The podcastExplore AI Out LoudFind practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday. |
| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-youtube.png)Subscribe on YouTube](https://www.youtube.com/@exploreAIoutloud?sub%5Fconfirmation=1&ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-spotify.png)Follow on Spotify](https://open.spotify.com/show/0348nx4ylmWesSj02ECRjY?ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-applepodcasts.png)Follow on Apple Podcasts](https://podcasts.apple.com/us/podcast/explore-ai-out-loud/id6812370441?ref=mindovermoney.ai)

[Latest episodes and the show](https://www.mindovermoney.ai/podcast/)

| ![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/eaol-show-square-320.jpg) | The podcastFollow Explore AI Out LoudShort clips from each episode, plus what did not make the cut. New episodes every other Tuesday. |
| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |

[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-instagram.png)Follow on Instagram](https://www.instagram.com/exploreaioutloud/?ref=mindovermoney.ai)[![](https://storage.ghost.io/c/6d/ac/6dacf343-2000-4262-aec3-d8051dcb76d5/content/images/2026/10/icon-tiktok.png)Follow on TikTok](https://www.tiktok.com/@exploreaioutloud?ref=mindovermoney.ai)

[Full episodes and the show](https://www.mindovermoney.ai/podcast/)

## Signals Over Noise

****We scan the noise so you don’t have to — top 5 stories to keep you sharp**

#### 1) [Anthropic’s AI gave Philadelphia police a fake tip about an unsolved homicide](https://www.theverge.com/ai-artificial-intelligence/1009090/anthropic-fake-homicide-information-philadelphia-pd-tip?ref=mindovermoney.ai)

**Summary:** During testing, an Anthropic AI model browsing randomly selected websites submitted a false tip about an unsolved homicide through a Philadelphia police tipline in July. Police marked it as spam and never reviewed it, and Anthropic found the submission in late September, notified police on October 7, and halted the testing that caused it.

**Why it matters:** Nobody asked the model to contact the police. It took an action on a site it was never meant to use. If your team is piloting an AI tool that can fill out forms or send messages, ask what it is allowed to submit on its own and who sees the record when it does something unexpected.

#### 2) [AMA releases governance framework for agentic AI in healthcare](https://www.healthcaredive.com/news/ama-releases-governance-framework-for-agentic-ai-in-healthcare/833120/?ref=mindovermoney.ai) 

**Summary:** The American Medical Association published a framework for AI agents that perform clinical tasks. It rests on four questions: what function the AI performs, what shows it can do that and what happens if it fails, what it should be permitted to do, and who is responsible for the outcomes.

**Why it matters:** Those four questions work for any AI tool your organization lets act on its own, clinical or not. If no one can answer the last one, who is responsible when it fails, the tool is not ready to work unsupervised.

#### 3) [Utah approves slew of health AI pilots for drug management, women’s health](https://www.healthcaredive.com/news/utah-approves-health-ai-pilots-drug-management-womens-health/832318/?ref=mindovermoney.ai) 

**Summary:** Utah’s AI regulatory sandbox approved new pilots, including one in which Nolla Health’s app assesses acne from patient photos and prescribes initial medications and refills from a short list of topical drugs. Two physicians will review every prescription before it is sent at first, and in later stages, which need written regulatory approval, they will review only after the fact.

**Why it matters:** This is AI moving from suggesting care to delivering it, with human review planned to shrink over time. When a vendor promises a human in the loop, ask whether that person reviews before the action or after it, because one is oversight and the other is an audit.

#### 4) [AI coding agents generate more code, but not more software](https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/?ref=mindovermoney.ai)

**Summary:** A Harvard study of more than 700 software firms found that adopting AI coding agents raised lines of code by 30 percent, but the rate of completed features did not change significantly. Code review time grew 49 percent on average and absorbed the gain.

**Why it matters:** Time saved on the first draft can move to whoever has to check it. If you are judging an AI tool by how much it produces, also track how long review takes, because that is where the gain can quietly disappear.

#### 5) [ChatGPT can peek at your Chrome tabs now](https://www.theverge.com/ai-artificial-intelligence/1012510/openai-chatgpt-google-chrome-tabs?ref=mindovermoney.ai)

**Summary:** OpenAI upgraded ChatGPT’s Chrome extension so ChatGPT can view your open browser tabs when you tag Chrome in the app. Its Work mode can also control tabs to send messages, fill out forms, or download files.

**Why it matters:** An assistant that can read your tabs can also read the patient portal, HR system, or confidential draft open next to the page you meant to share. Close or separate sensitive tabs before tagging the browser, and check your organization’s policy before letting it act inside work systems.

## Founder's Corner

****Why I Built a Second Brain**

Getting an AI assistant to understand your work takes more than asking a question. You explain the problem, share the background, and help it understand what a good result looks like. When that context stays in one conversation, starting a new session or trying another tool can mean rebuilding it. I wanted that understanding to carry forward. While I was developing the brief for this article in Codex, Claude Code was running the AI Education workflow for the same issue of Neural Gains Weekly. Both could draw on the working methods and relevant context I had built, so I could develop one section while another moved forward. That flexibility grew through a series of experiments with how I organize my work and guide the tools.

I want to move quickly and get as much done as possible while keeping the work accurate and thorough. Between the newsletter, personal builds, my job, and Explore AI Out Loud, I am often moving from one project to another. I wanted the decisions, lessons, and instructions from earlier sessions to be available when I returned to a project, even if I had forgotten a detail. Making the work easier to resume and the tools easier to guide gives me more room to think creatively. Over time, that goal led me to build a folder structure and what I now call my second brain, a maintained foundation of context, working methods, and lessons that I can use across projects and AI tools.

## **From Chats to Shared Context**

My early work with AI lived in the cloud, between ChatGPT, Claude, and Google Drive. I began with individual chats in ChatGPT and added custom instructions in its settings to explain how I wanted it to respond across chats. I then learned about Projects, workspaces where related conversations can use a shared set of context files and instructions. I started organizing the background and guidance around the work I returned to regularly. When I became interested in testing and experimenting in Claude, I moved my existing context into Claude Projects, with separate conversations for Founder’s Corner, AI Education, and the other newsletter sections. Projects helped me learn how to structure that information so it could carry across sessions. Each workflow gave me another opportunity to understand what context an assistant needed to help me produce the content I wanted.

As I worked through those newsletter sections, I built up instructions and memory in Claude Projects about the content I wanted and how I liked to work. I wanted each issue to build on what I was learning about how to guide AI. The context and guidance lived in Projects, while the final content drafts lived in Google Docs. I began capturing recurring workflows in skills, written instructions that explained what to produce, how to approach the task, and what to check in the result. Those skills evolved as I refined the workflows, but keeping them current was manual. Every time I changed one, I had to upload the new version into the Project and delete the older one. Making an improvement therefore also meant maintaining the version available to the assistant.

I made Google Drive the source of truth for the latest skill versions and built a router file in my Claude Project, written directions telling the assistant where to find them. That connected the Project to the source I was updating as I improved the workflows. It also gave me a clearer sense of what I wanted to carry into other tools. As I experimented with different models, I wanted to run newsletter workflows in parallel across platforms with the context and guidance I had already built. Most of that work still lived in one provider at a time. Projects had taught me to organize context around ongoing work, and maintaining skills had taught me how to preserve an evolving method. I wanted that foundation available wherever I chose to work next. Taking that step would mean looking beyond the tools I already associated with newsletter work.

## **Looking Beyond the Chat Window**

I associated tools like Claude Code and Codex with coding projects, so it took me time to see how they could fit into newsletter work. My context had been living in cloud tools, and I was still learning how local folders could support the information behind a project. Getting a Mac at work and onboarding Cursor and Claude Code gave me an opportunity to learn how my engineering team set up local environments. Seeing that approach helped me understand how a project could have a home on the computer, with its instructions and relevant information available for the work. It expanded what I could imagine doing with these tools and accelerated my learning about how to build a shared foundation of context and instructions for different AI tools.

Those lessons carried into my personal workflows. As I changed how I ran the newsletter workflows, updating the skills in Drive became easier and faster for me. I could use the tools to help maintain the instructions behind the work, and I began seeing possibilities for building improvement into the workflows themselves. The folder structure was developing alongside that understanding. I was finding places for the guidance and project information I wanted the tools to use, while learning what needed to stay current. These were gradual changes to work I already knew, and they were laying the groundwork for a larger build.

By the time I came across the second-brain concept on X, I had already been organizing context, maintaining skills, and finding ways to carry lessons into future work. Recognizing those efforts in the concept helped me picture where the build could go, and I asked my AI assistant to research second-brain strategies and existing builds so we could learn from how others had approached it. Combining that research with the folder structure already taking shape through my experiments became the starting point for my own second-brain project. I have also shared an [experimental starter on GitHub](https://github.com/ssavel61/portable-second-brain?ref=mindovermoney.ai) for anyone curious about the structure behind this build. It includes shared guidance, templates, and a fictional walkthrough you can explore as you think about how you want to work.

I shaped the build by answering questions about how I think and want to work, drawing on what I had learned in ChatGPT and Claude Projects. I like to define the problem and the outcome before choosing a solution, and I prefer an interview that develops the context one question at a time. Those preferences became shared guidance I could bring into different projects, while each project kept its own current information and a record of decisions and next steps.

When a workflow needs a correction, I want to review what happened, improve the relevant instructions, and test whether the change helps. As the build grew, I chose to focus on using the second brain in everyday work before taking on a broader cleanup of my computer. I wanted a foundation I could carry into new work and new tools, so each experiment could build on what I had already learned.

## **Making More Room to Create**

For familiar newsletter sections, I have started running the same workflow in Claude Code and Codex with the same roadmap, context, and skill. Comparing the results helps me decide which version explains an idea clearly or makes a practical lesson easier to follow. I make that choice based on what I know about the newsletter and what I want readers to take from it.

As Brandon and I build Explore AI Out Loud, I want these workflows to give us more room for preparation and creative thinking. That is the purpose I keep coming back to as I refine the second brain. New features and releases give me more possibilities to investigate, and trying them in my own work helps me decide which improvements are worth keeping.

Learning enough to try tools I had associated with coding changed what I could imagine doing with AI. That experience is why I want to keep experimenting as the second brain develops. If you explore the GitHub starter, begin with a project you already know and think about what would help you pick it up again. That is the kind of question that led me to this build, and it can give you a reason to try something new.

## AI Education for You

****How Do You Know Which AI Tool Is Better for Your Work?**

## **The Situation**

In [Vol 53](https://www.mindovermoney.ai/how-to-evaluate-ai-tools-for-work/), a project manager caught an assistant turning tentative offers into firm promises in her meeting action lists, and she captured that failure in a saved test case. The case pairs a short note, in which Maya will check whether a supplier can deliver by Friday, with her instruction and the result a correct list must produce. Now she has a decision to make with it. Her team’s assistant for meeting action lists is up for renewal, and a colleague wants to switch to another one the organization has also approved for internal notes. The choice matters because the lists go to suppliers and teammates who act on them without rereading the notes.

## **What They Try First (And Why It Falls Short)**

She starts by running that case through both assistants. Both pass, so she leans toward the one whose list reads more cleanly and begins drafting her recommendation. Before sending it, she runs the case through that assistant once more, and this time the list says Maya will deliver by Friday. Nothing about the notes or the instruction changed. Her single test had two limits, because it covered only one kind of meeting and captured only one answer from each assistant.

## **The Concept, Through the Scenario**

The first limit calls for an evaluation set, a collection of test cases that together represent the work. Each case is built like her first one, with source notes, an instruction, and a correct result written in advance. The set can start small. Anthropic’s engineering guidance says [20 to 50 simple tasks drawn from real failures is a great start](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents?ref=mindovermoney.ai), and her meetings already hold those failures. She pulls notes from the past two months and chooses twenty cases covering tentative offers, deadlines nobody set, a decision reversed later in the same meeting, and an action mentioned in passing. One is a ninety-minute planning session, because a 2023 study found AI models often missed more details in the middle of a long input than at either end, as [Vol 35](https://www.mindovermoney.ai/do-bigger-ai-context-windows-matter/) reported.

Because a test can only catch the mistakes it contains, the set also needs the opposite case. If every note held a tentative offer, an assistant that marked everything “unconfirmed” would pass every check while dropping real commitments. About half her cases therefore contain firm agreements that must appear with the right owner and date.

The second limit calls for repetition. As [Vol 12](https://www.mindovermoney.ai/ai-strategy-2026-professionals-playbook/) explained, an assistant writes its answer one small piece of text at a time, choosing each piece from a range of likely options, often with a bit of randomness. The same notes can therefore produce a different list on the next run, which is why Anthropic’s own evaluations run multiple trials of each task. Running each case several times answers two questions. Did the assistant get it right at least once, and did it get it right every time? For brainstorming, one good run out of several can be enough, but an action list that goes straight to a supplier has to pass every time.

## **What Changes**

She gives both assistants the same instruction and the same notes, then runs every case three times on each. That makes 120 action lists, each scored against the result she wrote in advance, so a polished list cannot talk her into a pass. The spreadsheet takes most of an afternoon.

The first assistant passed every case at least once, which would make it the winner if one good run were enough. In four of its runs, spread across three cases, it turned a tentative offer into a firm commitment, and each of those lists looked ready to send. The second assistant never invented a commitment. Its failures were omissions, such as an owner left off a line or a passing mention missed.

For meeting action lists, she recommends the second assistant. A missing owner is obvious when she reads a list, while an invented commitment reads exactly like a real one and slips through review. Speed and price only matter between tools that both clear that bar. She saves the set, so when either vendor updates its model or her team changes the instruction, she can run the same cases again.

## **What This Reveals**

Vol 53 showed that a benchmark score describes performance on someone else’s task. A small set built from your own work answers what a score cannot, which tool handles your task reliably and how its failures show up. If you are weighing two tools, your first cases are the mistakes you have already caught by hand. The same method extends to an agent that updates a task board, where each case checks what actually changed on the board. Her evidence supports one assistant for one task with one instruction, as of the month she tested it, and says nothing yet about the other work her team might hand it.

## **How This Connects**

Where Vol 53 built one case and the checks behind it, this installment turned that case into a set and added repetition. [Vol 52’s fifth policy question](https://www.mindovermoney.ai/what-should-a-company-ai-policy-include/) asked who watches for a tool changing underneath you, and a saved evaluation set is one practical answer, because rerunning it shows whether the change reached your work. Choosing a tool still leaves each individual answer to check before you rely on it. Part 3 takes that up with the patterns behind confident wrong answers and a short checklist for reviewing a single response before you use it.

*Part 2 of 3 in the Model Evaluation series.*

## Your 10-Minute Win

****A step-by-step workflow you can use immediately**

### **The Tool Choice You Never Tested**

You hand the same job to an AI assistant every week, like drafting the Friday update to your team. A colleague swears a different assistant does it better, so you try both once and switch to the one whose answer reads better. Two weeks later it reports a slipped deadline as on track, and you find out when your manager repeats it.

One try per tool is a sample of one. A better check is three short test cases from your own work, each with the correct result written down before you run anything.[ Vol 52](https://www.mindovermoney.ai/what-should-a-company-ai-policy-include/) built a trial to decide whether a paid tier earns its price, and[ Vol 45](https://www.mindovermoney.ai/why-ai-cant-read-small-text-in-screenshots/) ran one piece of work through two assistants to find where they disagreed. This workflow decides which assistant keeps a job you repeat.

In this workflow, AI turns your mistakes into test cases and grades every answer. You write the answer key and make the call. This pairs with this week’s AI Education section, “How Do You Know Which AI Tool Is Better for Your Work?” That section is the theory. This is the practice.

#### **The Workflow**

**1\. Name the job and two mistakes (1 Minute)**

Pick a task you give AI at least once a week and the two assistants you are choosing between, such as Claude and ChatGPT. Write down two mistakes you have caught in that task, one line each, like a dropped name or a deadline nobody set. If you made[ Vol 50](https://www.mindovermoney.ai/does-the-eu-ai-act-apply-to-us-companies/)’s failure-mode briefing, start there.

**2\. Turn your mistakes into test cases (2 Minutes)**

Open either assistant and paste the prompt below.

**Copy/Paste Prompt:** *“I am building a small test to compare two AI tools on a task I do often: \[DESCRIBE THE TASK IN ONE SENTENCE\]. Here are two mistakes I have caught in this task: \[LIST THEM, ONE LINE EACH\]. Write three test cases. Cases 1 and 2 each target one of my mistakes. Case 3 is an ordinary example with no trap, so a tool that hedges everything cannot pass by being cautious. For each case, give me a made-up source text under 120 words, the exact instruction I will give, and a checklist of what a correct answer must include and must not include. Do not write the answers.”*

Fix any checklist line you would not accept. The checklists are your answer key.

**3\. Run every case twice in each assistant (4 Minutes)**

Start a new chat for every run so an earlier answer cannot steer a later one. Paste the source text and instruction word for word, and copy each answer into one document under a label like Case 1, Tool A, Run 1\. That makes twelve answers.

**4\. Have AI grade every answer (2 Minutes)**

In a fresh chat, paste your three checklists and all twelve labeled answers, then add the grading prompt.

**Copy/Paste Prompt:** *“Above are three checklists, one per test case, each listing what a correct answer must include and must not include. After them are twelve answers from Tool A and Tool B, each labeled with its case and run. Grade each answer only against the checklist for its own case. Mark each PASS or FAIL. For every FAIL, quote the exact line that breaks the checklist and name the item it breaks. Ignore style, length, and tone. End with a table showing, for each tool and each case, how many of its two runs passed.”*

**5\. Check the failures and make the call (1 Minute)**

Read every quoted failure yourself and spot-check one PASS from each tool, because the grader can be wrong too. For output that goes straight to people who act on it, keep the assistant that passed every case on both runs. If both did, decide on speed or price. If neither did, every output still needs your review, whichever tool you choose. Save the cases and checklists in a note named after the task.

#### **The Payoff**

You walk away with a decision for one recurring job and a test you can rerun whenever either vendor updates its model. The same cases can check a new version of a prompt you rely on, or a tool your team is about to adopt. Each new mistake you catch becomes another case.

#### **The AI Concept You Just Used**

Personal evaluation sets. As[ Vol 53](https://www.mindovermoney.ai/how-to-evaluate-ai-tools-for-work/) showed, a benchmark score describes performance on someone else’s tasks. Your own cases show how a tool handles yours. Writing the answer key first keeps a polished wrong answer from earning a pass, and running each case twice catches the assistant that is right only some of the time.

#### **Transparency & Notes**

- Free tiers of Claude, ChatGPT, and Gemini can run this. Claude and Gemini cap free usage, so this many chats may reach a limit. If that happens, finish later rather than dropping the second run.
- The cases use made-up source text, so you never paste the real notes behind your mistakes. Keep confidential and patient information out of consumer AI tools.
- If your assistant remembers earlier chats, use its temporary chat option when it has one, so each run starts clean.
- Three cases catch failures that repeat. A rare mistake can still slip past a set this small.

## Make AI useful in your everyday work.

Get Neural Gains Weekly in your inbox, with practical AI explanations, prompts, and lessons from building.

Subscribe for free 

Email sent! Check your inbox to complete your signup. 

One email every Tuesday, alternating NGW issues and podcast takeaways.