This is Masaki Hirokawa, artist and indie app developer running Gemini Lab.
On a quiet Monday evening between holidays, almost on a whim, I zipped up the entire dev_notes/ folder from my Dropbox. Twelve years of development notes since I went independent in 2014. Release diaries, monthly AdMob reflections, late-night Crashlytics scribbles, App Store rejection histories, self-directed code review memos, store description drafts that never shipped. All of it. 38 MB compressed, roughly 700K tokens when counted as text — a near-perfect fit for Gemini 2.5 Pro's 1M context window.
Why bother? I had been writing several long-context articles around Google I/O 2026, and an uncomfortable feeling kept surfacing: I had never seriously used 1M tokens on something that mattered to me. I had played with sample data, of course. But I had never handed over 700K tokens of my own life as a developer. It reminded me of something an older artist told me when I was seventeen: "If you really want to understand something, do it once all the way through first."
The short version: when I stopped asking for summaries and started asking my past self questions through Gemini, this turned into a different kind of tool for solo developers. Here are the four things that quietly shook me that night.
Realization 1: My 12-years-ago self had already written the answer to today's problem
The first thing I tried was a small technical decision I had been stuck on all week. I was revisiting the ad-load-failure retry strategy for Beautiful HD Wallpapers v2.1.0 and could not decide whether the initial backoff should be 500 ms or 1000 ms.
I asked Gemini: "In my dev_notes/, what conclusions have I previously reached about ad retry intervals? Please cite the date and the reasoning."
What came back was a memo from September 2018. During a period when AdMob fill rate had collapsed, I had written: "initial delay 750 ms, max 3 retries, ±150 ms jitter — fill_rate recovered from 81% to 89%." I had completely forgotten. Gemini also pointed out that I had revisited the same debate in 2021 and 2023, landing on similar conclusions all three times.
What is interesting here is that Gemini did not produce the answer. My 8-year-old self produced it; Gemini merely fetched it. But the ability to pull the right passage out of 700K tokens in three seconds nudges the quality of solo-developer decisions up one notch. The most accurate description I can give: this is a tool that quietly reminds you that your past self was smarter than you remember.
Realization 2: I have been repeating the same mistakes roughly every three years
Half jokingly, I asked: "Across the last 12 years, can you identify three patterns of mistakes I have repeated? Please name three." The reply landed harder than I expected.
The first: the habit of overreacting to revenue on the day after a feature launch. In 2017, 2020, and 2023, I had written panicked late-night memos when day-three revenue did not move. Gemini laid them out as a timeline and noted, dryly: "Judging on day three has consistently led to regret in subsequent weeks."
The second: a tendency to postpone iOS StoreKit migrations until the last possible moment. Notes about StoreKit 2 migration appeared sporadically from 2022 onward, and yet here I am in May 2026 doing the migration across four apps at once. Since this is literally the work in front of me this week, I had to laugh.
The third: after international recognition of my art work, I push application changes too aggressively the following week. The week after the A'Design Award gold in 2023 and the week after the Luxembourg Art Prize in 2025, I had drawn up unreasonably ambitious roadmaps in both cases. This one I had genuinely never noticed.
Gemini is not clever here. What is new is simply that I now have a conversational partner who can lay out 12 years of my own behavior as a timeline. And once those three patterns were named, by the next morning I had the will to design small countermeasures so I do not create a fourth.
Realization 3: My app revenue curve and my art production periods are correlated
The third question put words to something I had vaguely felt for years. "If I lay the monthly revenue curve of my apps since 2014 alongside the completion dates of my artworks, do you see any relationship?"
Gemini replied: "I see a tendency for new art series to begin shortly after quarters in which app revenue growth has stalled." Specifically: after the Q2 2017 plateau, the photo-collage series Layers of Memory began. After the Q3 2020 plateau, I started experimenting with three-dimensional pieces. After the Q1 2024 plateau, I began preparing the works that would later show at LA ART SHOW.
This is correlation, not causation. But it let me reframe something: when the app numbers stall, I am not escaping into art — I may be searching for the next subject in a different language. The "intuition arrived" feeling I had over Kichijoji Station in 2019, when I saw a halo of light in the sky, may be quietly arriving inside me during revenue plateaus too. If so, that is not a habit to scold — it is a habit to trust.
My grandfathers, both temple carpenters, used to say: "Keep your hands moving and the shape of the next job will eventually appear in them." My 700K tokens of personal notes proved that sentence, twelve years later, in a small way.
Realization 4: I could finally see what to stop doing for the next 12 years
For the final question I asked: "If there were three things I could safely stop doing over the next 12 years, what would you recommend, based on past notes?"
Three answers came back: (1) the launch-day Twitter ego search, (2) the quarterly fatigue of swapping App Store screenshots, and (3) the end-of-year learning roadmap I set every January and never complete. Each one was supported by Gemini's accurate quotations of notes where I had written, over the past 12 years, that I intended to stop — and then didn't.
I do not believe AI can make me stop something I have failed to stop for 12 years, in three seconds. But being shown that my past self has written the same intention 12 separate times does plant the small, useful wish to not write it a 13th time. For the first time, what I should stop doing over the next 12 years rose up out of my own notes rather than out of a productivity book. That direction — what to subtract — feels more important to me right now than what to add.
Closing — Carrying 700K tokens of your past self with you
Gemini 2.5 Pro's 1M context is usually discussed in terms of benchmarks and code review. But for a solo developer, its real value might live somewhere else. Carrying your past with you as an always-available conversational partner is what I quietly felt the night I handed over twelve years of notes.
You do not need 12 years to try this. A single year of release diaries, six months of Crashlytics notes, or three months of revenue records will do. Rather than asking for summaries, try asking your past self today's question. Meeting the version of yourself who has repeated the same mistake five times is a bit embarrassing, but I have come to feel that once every three years, that embarrassment is what becomes the next step.
For now, I plan to make the last Monday evening of each month a small ritual — feeding that month's dev_notes/ into a 1M context window and spending thirty minutes asking my past self questions. If the next twelve years can be walked while borrowing the voice of who I used to be, I would consider that a quiet privilege.
Thank you for reading to the end.