A product manager at a 30-person software company opens the Google AI chatbot, pastes in an angry customer email and types: “How should I reply to this?” Back comes three paragraphs of smooth, agreeable nothing. She rewrites it herself in eleven minutes and closes the tab.
That scene plays out in thousands of offices every week. The tool isn’t the problem. The request is.
Gemini, Google’s chatbot, shows up in more places than most people realise: the standalone app, AI Overviews on the search results page, a side panel inside Docs and Gmail, and a shortcut on recent Android phones. Each one behaves a little differently. Knowing the rough edges before you build a habit around it saves real frustration, and a short read on what the Google AI chatbot does well and where it lets you down is a sensible place to start. After that, this six-step process turns a pleasant answer machine into something you’d hand actual work to.
Step 1: Pick the right surface before you type a word
Same model, four front doors, and they are not interchangeable.
- gemini.google.com or the Gemini app: file uploads, long documents, saved custom instructions. Best for anything involving your own material.
- AI Overviews in Search: fast, links to sources, shallow. Fine for “what’s the difference between X and Y”, useless for drafting.
- Workspace side panel: sees the Doc, Sheet or email you have open. The right call when the work already lives in Google’s apps.
- Android or Chrome shortcut: convenience only. Don’t run a serious task through it.
Most people grab whichever is nearest, get a mediocre answer, and conclude the model is mediocre. Walking twenty seconds to the app with the file upload button changes the result more than any prompt trick.
Step 2: Write a brief, not a question
Compare two requests sent to the same chatbot on the same afternoon.
“Write a post about our new coffee subscription.”
“You’re writing for a specialty coffee roaster with 1,800 Instagram followers, mostly home brewers aged 25-40. Write a 120-word caption launching a £14/month subscription. No exclamation marks, no ‘elevate’. One concrete detail about the beans: washed Ethiopian, roasted every Tuesday. End with a single question.”
The second takes ninety seconds to write and produces something usable on the first pass. The pattern never changes: role, task, audience, constraints, format. Skip the constraints and the model fills the gap with the blandest average of the internet. There’s a fuller breakdown in briefing an AI chat like a new colleague, including the phrasing that tends to work.
Step 3: Paste the raw material, not a description of it
Describing your data costs accuracy and gains nothing. Paste the thing.
A freelance illustrator raising her rates fed in her last three client emails plus a rough list of twelve projects with hours attached. The instruction: “These are real projects. Identify the three that took the most hours relative to the fee, and draft a rate increase email for those clients only.” That produced a specific, defensible email in one pass. Asking “how do I raise my rates?” would have produced a listicle.
Gemini handles long context reasonably well, though it drifts past a certain length. If you’re pasting forty pages, re-anchor it halfway with “reminder: the goal is X”. The surrounding loop matters too, and there’s a five-step workflow for using an AI chatbot online that covers intake, verification and handoff.
Step 4: Ask for the skeleton before the flesh
The single most useful line I’ve found:
“Before you write anything, give me an outline of no more than five sections. For each, state the one point it makes and what evidence it needs. Tell me what you’d cut.”
This gives you a cheap place to disagree. Restructuring five bullets takes a minute; rewriting 900 words takes twenty. It also surfaces the model’s assumptions while they’re still cheap to correct.
Step 5: Make it argue with itself
First drafts from any chatbot skew agreeable. It writes what you seem to want. Force a second pass:
“Now argue against your own answer. What would a sceptical finance director pick apart? What’s the weakest claim, and what number would I need to prove it?”
On a budget proposal, this surfaced that the projected 22% saving assumed no rise in support tickets, an assumption the original draft had buried in a clause. A human reviewer catches that on a good day and misses it on a busy one.
Step 6: Bank the prompt that worked
A prompt that worked once is an asset. I keep one doc with about thirty of them, each named for the job: churn email, job ad v2, meeting notes to actions, competitor teardown. New task, copy the nearest one, swap the variables.
If your team keeps repeating the same request, that’s a sign to consider building a customer-facing bot that survives real customers rather than asking five people to paste the same prompt every morning.
Where the Google AI chatbot still trips you up
The failure modes are predictable after a few months of use. Watch for these:
- Confident numbers. It will produce a statistic with no source. Treat every figure as unverified until you find the primary source.
- Source drift. When it cites the web, the link sometimes supports a weaker claim than the sentence built on it.
- Long-chat drift. Deep into a thread, it starts answering a slightly different question than the one you asked.
- Hedging. Ask for a recommendation and you get a menu. Reply “pick one and defend it” and it usually will.
- Formatting bias. Left alone it reaches for bullet lists and bold text. Request three tight paragraphs and the prose improves.
None of these are dealbreakers. They’re reasons to keep a human in the loop for anything that leaves your desk. Rebuilding each request from scratch is the common trap here, and there’s a good walkthrough on going from vague answers to useful results aimed squarely at that habit.
A full example, start to finish
A four-person physiotherapy clinic wanted to cut no-shows, running at about 14% of appointments.
The brief: “You’re a practice manager. I’m pasting the last 60 appointment notes where the patient didn’t show. Group them, name the two most common patterns, and draft a reminder text under 40 words.” Pasted 60 rows. Asked for the outline first, which produced three candidate patterns. The red-team pass killed one as too small a sample to act on.
Result: two reminder texts, one for first-time patients and one for people booking early-morning slots, both under 40 words. Total time, twelve minutes including reading. The clinic tested them for a month and no-shows fell to roughly 9%.
The 20-minute test
Pick one task you do every week: the Monday status update, the quote email, the job posting you keep postponing. Time yourself doing it the normal way. Then run it through the six steps and time that.
Some tasks won’t be worth it. Anything where you already know the answer, or where the value sits in your judgement rather than the words, tends to come out even. The ones where you’re staring at a blank page with a pile of messy inputs nearby are where the gap shows up, usually three to five times faster with a comparable result.
Run that test on two tasks and you’ll know whether the Google AI chatbot earns a place in your week. Skip it and you’ll keep getting smooth, agreeable nothing.