I'm Masaki Hirokawa — an artist and a solo app developer, and the person running Gemini Lab.
Lately, readers keep asking me the same question: "Which Gemini should I actually use?" Between 2.5 Pro, 2.5 Flash, 3.1 Pro, Deep Research, NotebookLM, and Google AI Studio, the menu has roughly doubled in the last six months. I genuinely hesitate every time myself.
So today, instead of a how-to, let me write something a bit more personal: a day in the life of someone who builds and runs apps on their own, and where each Gemini variant actually shows up in that day. This is a snapshot as of April 2026 — a few months from now, it'll probably look different, and I'll tell you about that then.
6:30 a.m. — The first reach is almost always for Flash
While the coffee is brewing, the first thing I open is Gemini 2.5 Flash. The reason is simple: at this hour, my brain isn't really online yet, so what I need is speed and lightness, not depth.
Concretely, I'll drop three to five overseas articles I'd bookmarked the night before and ask, "Summarize each of these in three lines." Pro could do the same work, but at this hour, throughput matters more than precision. Flash answers fast, and my half-awake brain skims at roughly the same rhythm.
What I don't touch at this hour: code review, release note translation, anything tax-related. Decisions I make before my brain wakes up tend to come back and bite me.
10:00 a.m. — For deep implementation work, it's 2.5 Pro or 3.1 Pro
When the main task of the day is "ship a new feature" or "chase down a nasty bug," I switch to 2.5 Pro or 3.1 Pro. As of April, my rough feel is:
- 2.5 Pro is the best at picking up context from my existing codebase and telling me, "If you change this here, these other places will feel it."
- 3.1 Pro more often surfaces an option I hadn't considered — especially on larger, cross-file design decisions.
Both are strong, so honestly I pick based on the "mood" of the task. Small fixes touching three files or fewer go to 2.5 Pro; rethinking an architecture goes to 3.1 Pro.
One thing I've learned: when I reuse a long system prompt, 3.1 Pro behaves more consistently. With 2.5 Pro, long contexts occasionally produce "over-committed" suggestions — which is fun on good days, and a reason my conversations wander on bad ones.
Early afternoon — Deep Research is like outsourcing the reading
Right after lunch, roughly 1:00 to 2:00 p.m., is the foggiest part of my day. That also happens to be when humans are worst at doing background research.
These days, I've started handing long prompts to Gemini Deep Research and then going for a walk. App-category research, competitor review patterns, pricing and monetization trends across Japan and overseas — anything I want to sit with for 30 to 60 minutes, I hand over and leave.
What I love about Deep Research is that it pulls in secondary sources I would never reach on my own — forums and niche blogs in English especially outrun any attempt to follow them solo. By the time I come back from the walk and read the summary, my head has cleared and I can move straight into decisions.
I still verify every concrete fact myself, though. Deep Research is an investigator, not a witness. That framing feels like the right distance for me.
Early evening — NotebookLM to make sense of "my own week"
By late afternoon, my coding energy is depleted but admin work fits fine. This is where NotebookLM earns its place.
I'll pile a week's worth of my own notes — conversations, half-formed ideas, memos — into a single notebook and ask, "What three themes kept coming back this week?" or "What discoveries in these notes haven't made it into a published article yet?"
The interesting part is that NotebookLM only reads what I give it, so it doesn't drift into generic internet takes on the topic. It feels a little like having someone quietly sitting just outside my own head, listening back to me. Working solo, you rarely get to see your own thinking from the outside, and this time of day has quietly become the anchor of my weekly rhythm.
Night — The first draft of a Gemini Lab article is still written by me
This might surprise some of you, but the first drafts of Gemini Lab articles — including this blog post — are still almost entirely written by hand.
The reason is that when someone finds the site from a search result, I want the text to carry the warmth of "whatever Masaki Hirokawa was actually thinking that day." If I hand the first draft to Gemini, the prose becomes cleaner, sure, but my hesitations and odd turns of phrase get averaged out.
In return, once I'm done writing, I hand the draft to Gemini 2.5 Pro and ask it to flag redundancy, awkward connectors, typos. It's very good at lifting a finished draft one notch higher. Since I started splitting the work this way, writing has become noticeably less painful.
Looking back at the day
Writing this out, I realize I'm quietly casting different members of the Gemini family into different roles across the day. Flash plays the impatient morning partner; Pro plays the reliable focus-hour coworker; Deep Research plays the colleague who goes off to read while I walk; NotebookLM plays the calm listener at dusk.
This isn't really a "which model is best" question. It's more about where the gaps are in my own day, and which model quietly fills each one.
If you're feeling lost in the current Gemini lineup, the single thing I'd suggest is this: write out your own day first. Coding, research, organizing, writing — once you see where each hour hurts, it becomes a lot clearer which model or feature is actually on your side.
If you want a more technical angle, you might also enjoy our Gemini 2.5 Flash vs Pro Selection Guide and Gemini Deep Research Guide.
I'm still figuring all of this out myself, so once I've lived with the lineup a while longer, I'll come back with an updated snapshot. Thank you, as always, for reading.