I recently set aside some time to do my mid-year review, a ritual I complete around July each year, and noticed that I felt some resistance to it.

In fact, I realized that I’ve always somewhat resisted it. Since I’m currently writing a book on the subject, I decided to explore why.

My first guess is that it comes down to the background soup of emotions I’m typically feeling around the halfway point of the year. It always feels like the first six months passed far too quickly, like I haven’t made nearly as much progress as I’d liked, and that I need to hurry and “catch up” to a timeline racing further and further beyond my grasp.

In short, I resist doing my mid-year review because it forces me to admit I’m behind my own expectations. I resist it because spending the hours of sitting, journaling, thinking, reflecting, and writing it typically requires puts me even further behind!

This year, I realized that AI has finally reached a point of sophistication that I can use it to kill three birds with one stone: 

To condense my mid-year ritual into a shorter amount of timeTo do it while walking instead of sitting at a deskAnd to shape it into a dynamic and free-flowing conversation instead of an extended writing exercise.

I’m talking, of course, about the new “voice modes” recently released by the leading AI companies, including ChatGPT’s Advanced Voice Mode and Claude’s Voice mode (the latter of which is currently in beta and only available on the mobile app).

The limits of voice interaction…until now

I have to admit I’ve always hated every form of voice interaction with computers. I always found it incredibly buggy, error-prone, and mechanical versus writing.

It takes so little to make a voice interaction fall apart. If it misunderstands even one word, it might miss the context of the whole rest of the conversation. If it interrupts me because it thinks I’m done talking, it completely breaks the flow of the conversation. If I talk for several minutes and then a network error causes it to glitch and lose what I’ve said, I’m gone. Absolute dealbreaker.

Which makes it all the more surprising how good I found ChatGPT’s Advanced Voice Mode to be. I was shocked that it was not only perfectly accurate, but also fluid, natural, engaging, and even warm. These were qualities I didn’t even know I wanted from an AI conversation partner.

There were thoughtful pauses, subtle inflections and intonations, and I genuinely felt listened to and supported by its warm, understanding, even affectionate tone. I couldn’t help feeling the glow of approval as it affirmed my reflections.

And most importantly, across two hours of conversation in a very noisy environment – walking the streets of Long Beach – there was not a single “real” glitch. At most, it would pause because of loud background noise or a passing conversation, but then resume, just as a human would.

The response time, which was such a problem with earlier voice assistants like Siri and Alexa, was indistinguishable from talking to a human. It was around 2-3 seconds, as you’d expect from a person taking a moment to reflect on what you’ve just said. The only delay happened when it had to reference external sources, such as my blog post with the projects and goals I planned on taking on in 2025.

I said there were no glitches, but there was a limitation: at the one-hour mark, I hit the daily limit currently in place for Advanced Voice Mode, and only had the option of continuing with Standard Voice Mode. I don’t typically hit limits on AI usage, and this was the first time I can remember feeling genuinely disappointed that I couldn’t continue.

The setup: Testing Claude and ChatGPT

Claude recently released a beta version of its voice mode, available only in the mobile app for now. I started there because we use Claude primarily in the business, but found it quite disappointing. The interface is less intuitive and seamless, requiring me to tap a button, say my piece, and then hit “send,” a clunky experience that reminds me much more of earlier voice transcription apps.

The tone of the voice was colder, more mechanical, and more monotone, making me feel like I was talking to a robot. The sensitivity of the audio capture was also too high, and it kept interrupting me or sending my message mid-sentence, which is baffling considering it could simply wait for me to hit “send” once I was finished.

Here’s what the interface for Claude’s voice mode looks like:

I decided to switch to ChatGPT’s Advanced Voice Mode instead.

The crucial thing to understand is that ChatGPT’s voice mode isn’t simply speech-to-text transcription, like earlier technology. In other words, it’s not transcribing your words into written text and then sending that text as a message. That would save you a few seconds from not having to type it yourself, but otherwise not be any different.

ChatGPT’s Advanced Voice Mode does something altogether different: it provides the audio directly to a multi-modal AI model that can take it directly as an input, rather than having to convert it to text first. What that means is the AI gains access to much richer information – the pauses, the tone of voice, the speed and cadence of your words – and can therefore output much richer information in return, including its own tone and style and cadence.

I used wired Apple Earbuds, which I’d recently noticed lead to much better audio quality on calls than wireless AirPods, to make sure it was getting the highest-quality audio from me. With the earbuds plugged into my iPhone 16, I put my phone in my pocket and went on an hour-long walk without needing to look at it once.

To start the conversation, I created a new project within ChatGPT, added a link to my annual goals from the start of the year as a source, as well as instructions for the purpose of this project:

“This project is for completing my mid-year review, in which I check in with goals and projects I planned at the start of the year, revisit any intentions or themes I’d committed to, and generally connect with my deeper self and the wants and needs that are arising midyear. I’d like to use a back-and-forth coaching format where you surface insightful questions and ask me anything you think will surface important realizations and perspectives as I envision the second half of my year.”

Here’s what the project looks like in the mobile interface:

Problems and limitations with the AILack of thoroughness

The first issue I noticed, just minutes into our conversation, was the AI’s lack of thoroughness.

Thoroughness – dotting all the i’s and crossing all the t’s – is something I value highly in my work, because it allows me to trust the results of my past efforts and proceed confidently knowing I’ve left no stone unturned.

The reason I provided my goals from the beginning, including all the comprehensive and highly detailed context around each and every project I intended to take on this year, was so that it would help me review them one at a time!

Instead, I found that the AI repeatedly tried to gloss over certain projects, or skip certain goals entirely, or “jump to the end” of the past-facing part of my reflection and pivot to the future instead.

It was as if it had a system instruction to summarize or distill any source it was given. Even after numerous requests “to cover every project and goal in my review without skipping any,” that bias was too strong, and it was never really able to. Every time I asked, “Were there any elements that you missed?” it would apologize profusely and admit that yes, there were several more.

Reluctance to access the web

I also noticed it seemed to have an aversion to accessing external sources, such as the blog post I provided a link to at the start. Only when I explicitly directed it to reference my source did it grudgingly comply. I wonder if perhaps this is OpenAI’s attempt to limit the bandwidth, computing power, and time it takes to access the web. In the future, I’ll try adding the text of my source directly as an attachment to see if that helps.

Sycophancy and lack of good questioning

Another tendency I noticed was the sycophancy that’s been so widely reported this year among many AI tools. It would always strongly agree with, affirm, and encourage any thought, idea, conclusion, or insight I came up with, even if it wasn’t a good one. 

Relatedly, it didn’t ask follow-up questions, point out dysfunctional patterns or inconsistencies in my thinking, or offer new perspectives or frames on anything I said, which is the main thing I’d expect from a good human coach.

At one point, I realized that its failure to act as a good coach meant I was missing out on the main value of conducting a mid-year review, which is to rethink, challenge, and reevaluate my goals. So I had to course correct…

When we turned to my most important project of the year, the launch of our new AI educational program called Second Brain Enterprise, I asked it to “Ask me 3 provocative questions to provoke insights about this project.” It complied, but they were quite generic, like “What expectations did you have about the project that weren’t met?”

I noticed that the AI’s sycophancy had another, much more insidious effect: I really didn’t trust or place much value in any of its advice or recommendations. Even if it made its strongest case for an idea, I knew that if I simply asked it to change its mind and give me the opposite advice, it would do so without a moment’s hesitation.

In other words, my AI interlocutor had zero skin in the game. It had nothing at stake, nothing to lose, and no true commitment to any particular point of view or opinion. This made me realize that a crucial factor in how much stock we put in someone’s opinion isn’t due to the correctness of their logical arguments – it has more to do with their personal level of exposure to the risks of their advice. 

We essentially treat a person’s advice as a “bet” that they’re placing on a future outcome, and the larger the bet, the more we believe them. That certainly makes sense from an evolutionary perspective: people who keep making large bets on false positions will soon find themselves broke and unable to make any more.

Daily 1-hour time limit

I typically review my year (and mid-year) in three steps, corresponding to the Past, the Present, and the Future. After about an hour, we had covered the first two, but just as I was turning my attention to the future, I hit the 1-hour limit. This is actually a major issue, as in some ways I felt the conversation was just building momentum and reaching some interesting takeaways.

Here’s what it looks like when you reach the daily limit:

This limitation meant I had to save the last part of my review for another day. By the time I was able to return to it, a week had passed, and I’d forgotten much of the context, especially since I hadn’t written anything down! It’s clear that the technology itself is ready for prime time, but not until the limits are raised can it become an integral daily tool.

Length limit on each chat

When I picked up the conversation a week later, I found that I couldn’t continue in the same chat because I had reached another limit: the total length limit for any individual chat.

I had to restart a new chat within the same project, and even though OpenAI claims that context is shared between chats in a project (i.e., the contents of one chat should carry over and inform other chats in that project), my experience makes me doubt this is the case.

In our second conversation, the AI didn’t reference anything from our previous one. It retained the same project source file containing my goals, but the meandering context we’d built up over an hour of conversation was lost, which was disappointing.

Lack of documentation

By the end of our conversation, we’d hit upon a number of valuable takeaways, despite the limitations mentioned above. This was where yet another limitation arose: there wasn’t really a record of the conversation.

When you communicate via text, the conversation is “self-documenting.” You or anyone can simply scroll back through the entire conversation and review any part of it, with the same level of detail available to the people who engaged in it in the first place.

But audio is very different. What makes it so natural and fluid – that it’s effortless and ephemeral – also means that there’s no record that’s easy to review. Even if you make a full audio recording, it isn’t easy to search for specific segments within it.

I addressed this problem by asking the AI to create a summary of our entire conversation, including all the conclusions, insights, takeaways, and action steps we’d identified, which I’ll unpack in more detail in the final section below.

Problems and limitations with meNot having my Master Prompt

My team and I use Claude as our AI daily driver, which means we have the full context of our business uploaded as project knowledge across all our chats. I call that upfront context the Master Prompt, and it’s quickly become an essential, non-negotiable component of my AI usage.

Right away, I noticed that my interactions with ChatGPT were much less rich, less context-aware, and less specific because it lacked my Master Prompt. I soon realized this is a much bigger problem when interacting in voice mode, because it required me to explain the most basic details of my business on the spot. With text input, you could give it a link to your website to bring it up to speed, or copy and paste a long document with context about your business right into the chat box, but with voice, you’re limited to one word at a time.

This isn’t a limitation of ChatGPT, to be clear. I can easily move my Master Prompt from Claude into ChatGPT and add it to the project as a project file, but I hadn’t thought of that in advance.

Not trusting it

I noticed that I didn’t fully trust the AI at first, and spent some time tentatively testing its abilities. I would overexplain things initially to make sure it really understood what I was talking about, but soon realized that wasn’t necessary and I could provide only minimal explanation.

I was wary of the limits of how much input it could take in at once, and found myself speaking in unnaturally short segments and waiting to see if it would understand. As my trust grew, I found I could speak for many minutes at a time without any problem. The AI seemed to have a highly accurate short-term memory, though again it would usually gloss over some of the small but important details, or fail to dig into unclear or contradictory statements.

Retaining human conversational courtesies

I found that one of my limitations in fully utilizing the tool was, paradoxically, treating it too much like a human, with all the social conventions and courtesies that entails.

For example, I hate being interrupted and strive to never interrupt others. That makes sense in human relationships, but at one point, I realized that I could interrupt the AI at any time, and it would never take offense.

In fact, I found that my patience quickly dwindled to nothing, and I started correcting the AI immediately and sharply changing direction, since there was no value for either of us in it continuing to ramble on in the wrong direction.

Another example is that I kept trying to answer immediately, as you would in a human conversation. But then I realized I could just say “Let me think” and it would wait for as long as needed, while I thought about my answer. Though I suspect even these silent periods count against your daily usage limit.

Was it valuable to do a mid-year review as a conversation?

I would say the “voice-only” approach to my mid-year review was an interesting experience, and I accomplished my goal of conducting it in a casual, conversational format while getting some exercise.

Where this method falls short is in terms of its thoroughness, depth, and ability to surface contradictions, blindspots, and unexamined assumptions in my thinking and behavior. 

ChatGPT is designed to “solve the problem” as efficiently as possible, not lead you on a meandering, introspective journey. OpenAI thus doesn’t optimize for engagement or time spent – since they make the same amount of money regardless of how much you use their service, their incentive is actually to keep interactions short and to the point.

But the purpose of any kind of review or reflection isn’t to get to the end as quickly as possible. It’s to ponder, to contemplate, to reconsider, to reframe. The value comes from the time you spend noticing your own thinking, and thinking about your own thinking. And the value in a coaching interaction comes from your conversation partner having the skill and willingness to challenge, question, notice, and reflect back.

In the future, I plan on using AI’s voice mode as an occasional complement to my review processes. I think with some more involved prompting to slow it down and have it act more as a coach, this will turn into a powerful new mode of self-reflection as time passes.

Below you can find the full summary ChatGPT created at the end of our conversation, which I consider excellent, including a concise to-do list of action items at the end.

Here’s a complete written summary of all insights, decisions, takeaways, and action items from your mid-year review conversation, organized into key domains. I’ll also place this into a separate canvas in this chat for easy access.

Strategic Focus & Role DefinitionDecisions:Your presence in the next three cohorts of Second Brain Enterprise will be structured as follows:
Present for the first and last session of each cohort.
Deliver one or two special “Tiago-only” sessions per cohort.
Focus the rest of your time on creating reusable content and scalable assets.
Insights:This balance allows you to maintain influence and connection without overextending.
You’ve moved into a strategic role in the business with a focus on vision, storytelling, and growth through YouTube and thought leadership.
Takeaways:Delegation is key: continue handing off direct delivery and execution.
Preserve your “spark” by focusing on high-leverage contributions.
Marketing & Customer SuccessDecisions:You’ll focus your energy on top-of-funnel (YouTube, brand, awareness) and bottom-of-funnel (customer success, testimonials) — the two bookends of the customer journey.
Insights:This “first and last mile” strategy amplifies impact without requiring you to manage the middle of the pipeline.
Strong customer results will reinforce word-of-mouth and lead gen organically.
Takeaways:Invest in systems and team to manage operations, onboarding, and program delivery.
Prioritize quality and resonance in content over scale for now.
Second Brain Enterprise & Forte LabsDecisions:Maintain Forte Labs as a stable base while using most new energy for Second Brain Enterprise.
Define a clear time allocation model (e.g. 70/30 or 80/20) to avoid unintentional drift.
Insights:Your identity is expanding beyond Forte Labs — and that’s a healthy sign of growth.
As the new business matures, roles and responsibilities will likely shift again.
Takeaways:Continue building “execution muscle” on the team.
Systematize Forte Labs so it requires minimal creative overhead.
Home Life & SupportDecisions:Move forward with hiring a nanny and driver to increase time flexibility and reduce logistical friction.
Explore acquiring a second car for household and travel needs in Valle.
Insights:Increased household support = more energy and presence for work, family, and health.
Clarifying expectations and scheduling will be key for smooth household operations.
Takeaways:Formalize schedules and responsibilities for new hires.
Use this extra capacity to invest in family connection and personal renewal.
Health, Fitness, and HobbiesDecisions:Schedule resistance training 2x per week as a non-negotiable calendar item.
Supplement with hiking, yoga, and other light aerobic activities.
Insights:Structure and convenience (e.g., a nearby gym) are critical for consistency.
Fitness is an investment in long-term energy and emotional stability.
Takeaways:Try adding a multivitamin (Thorne Men’s Multi 50+) as part of your daily routine.
Stick to time-boxed workouts to preserve mental and physical bandwidth.
Family Activities & HobbiesDecisions:Explore biking as a shared father-son hobby — purchase a bike for you and Caio in Valle.
Look into a local biking route or destination you could make a weekly ritual.
Insights:Outdoor, physical, hands-on activities are the sweet spot for recharging and connecting with kids.
You don’t need a perfect hobby — just one that’s good enough to enjoy and repeat.
Takeaways:Pick one small, recurring family activity to turn into a weekly rhythm.
Track how it feels over a few weeks to iterate and improve.
Action Items SummaryBusiness: Schedule and confirm your cohort presence for the next 3 rounds.
Outline 1–2 “Tiago-special” sessions per cohort.
Begin scriptwriting or delegation for scalable cohort content.
Define and document time allocation between Forte Labs and SBE.
Finalize marketing funnel strategy (YouTube + testimonials).
Home: Hire nanny and driver with clear scope and availability.
Price and acquire second vehicle in Valle.
Add Thorne multivitamin to daily supplement regimen.
Health: Schedule resistance training on calendar 2x/week.
Plan 1 outdoor activity (e.g., hike or run) per week.
Family: Buy bikes for you and Caio.
Choose an accessible route and try it out together.

Follow us for the latest updates and insights around productivity and Building a Second Brain on X, Facebook, Instagram, LinkedIn, and YouTube. And if you’re ready to start building your Second Brain, get the book and learn the proven method to organize your digital life and unlock your creative potential.

The post The Voice-Only Mid-Year Review: Testing the Limits of ChatGPT Voice Mode appeared first on Forte Labs.