I’ve used Gemini on everything from a 200-word product blurb to a 4,000-word report that needed fourteen sources checked. The pattern holds across all of it: Gemini is very good at getting you from nothing to something, and noticeably weaker at getting you from something to something good. Nearly every disappointment I’ve had came from asking it to do the second job.
That distinction matters more than any feature list. So here’s what the tool is actually like to write with, week after week.
Getting from a blank page to a rough draft
Ask for “a blog post about email marketing” and you’ll get 600 words of competent nothing. The output improves in proportion to what you put in. Give it a transcript, three pages of messy notes, the client’s brand guide and a clear line about who’s reading, and the first draft becomes something you can edit rather than something you have to replace.
Feed it raw material, not instructions
Uploading source documents is where Gemini pulls ahead of a plain chat box. I’ve dropped a 40-minute interview transcript into a single prompt and asked it to pull out every claim the interviewee made about pricing, with quotes, and it did it without losing the thread. Do that in a tool with a smaller context window and you end up chopping the transcript into pieces and stitching the results together yourself.
Ask for structure before prose
This one habit saves me the most time. Instead of requesting a full draft, I ask for three different outlines with genuinely different shapes: one argument-led, one list-led, one built around a case study. I pick the best, cut a section, then ask Gemini to expand one heading at a time. Section-by-section drafting stops the model from restating the same point three times with different nouns, which is the single most common failure I see in AI-drafted long-form.
Repurposing is where the time savings are real
Nobody enjoys turning a finished article into a LinkedIn post, a 120-word newsletter item, five FAQ answers and a subject line that doesn’t sound like a press release. Gemini handles that grind well.
The trick is adding a constraint that forces it to preserve your thinking. “Cut this to 120 words without losing the argument about refund timing” produces something usable. “Summarise this” produces a paragraph that says nothing in particular, very smoothly.
The failure modes you’ll notice by week two
The word tics
Left alone, Gemini reaches for delve, tapestry, landscape, “a testament to” and “in the realm of” with startling regularity. It also loves a tricolon and a rhetorical question to open a section. Ban them by name in the prompt. It works, and it takes ten seconds.
Invented specifics
Statistics, study names, quotes and dates can arrive fully formed and completely made up. Anything factual needs a second pass, and I don’t mean a glance. Gemini’s research strengths and its tendency to invent citations are worth understanding properly before you trust a single number it hands you.
Structural sameness
Give it no direction and every piece becomes intro, three points, recap. Ask explicitly for a narrative that opens mid-scene, or a piece that ends on an open question, or a comparison with no tidy verdict. It can do all of those. It just won’t volunteer them.
It agrees with you too easily
“Is this good?” gets you praise. “Which paragraph is doing no work, and where does the argument collapse?” gets you something you can use. Phrase the question as an attack and the answers get sharper.
Prompt habits that change the output
- Paste your own writing first. Two paragraphs of yours plus “match this voice” beats any adjective you could use to describe your tone.
- Set a word budget. “450 words” is a constraint. “Short and punchy” is a wish.
- Name the reader. “A procurement manager who has already read three of these” produces different copy to “business decision-makers”.
- Ban words in the same prompt you ask for the draft, not after.
- Request two versions with different structures, not two versions of the same thing.
- Ask it what it doesn’t know. “List anything here you’d need to verify” catches a surprising amount.
How it compares to the other assistants
For long, careful prose, Claude tends to hold an argument together more reliably across several thousand words. Gemini is faster, cheerfully integrated with Docs and Gmail, and better at chewing through large piles of source material. The comparison between Gemini and Claude comes down to whether you need a drafter or an editor. If you’re weighing up ChatGPT instead, where Gemini and ChatGPT each win is a fairer starting point than any benchmark chart.
Writing on your phone without wanting to throw it
Most of my Gemini use isn’t at a desk. It’s replying to a client email on a train, tidying a voice note into a paragraph, or turning a half-formed idea into something worth keeping. On iOS the app is polished but still trips over longer edits; what the iPhone app handles well and where it stumbles covers the details. The Android version is better at handing work off between apps, though it has its own rough edges, as this breakdown of Gemini on Android makes clear. Either way, short bursts work; a 1,500-word draft on a phone does not.
The most underrated use: editing your own work
Gemini writes a decent first draft. It reads a draft better than it writes one.
Paste 800 words of your own and ask for a paragraph-by-paragraph critique, flagging any sentence that repeats a point already made, any claim that needs a source, and any transition that jumps. I’ve had it catch the same weakness in my introductions three times running: I tend to explain what the article will cover before I’ve given anyone a reason to care.
Ask it to argue against your piece. Ask it what a sceptical reader would push back on in paragraph four. Ask it to cut 20% without losing any of the substance. These are jobs where its agreeableness works against it only if you let it, and where you keep full control of the final text.
Keep a short prompt library and stop starting from scratch
After a few months, the prompts that consistently work are the ones worth saving: a voice-matching prompt with two samples baked in, an outline prompt with three structure options, a line-editing prompt with your personal banned-word list, and a fact-check prompt that asks for sources you can actually open.
Ten saved prompts in a plain document will outperform any clever one-off instruction you invent on a Tuesday afternoon. Refine them when the output drifts, delete the ones that never earn their place, and treat the whole thing the way you’d treat a well-worn style guide. The model changes; your prompts shouldn’t have to.