Last month a freelance translator I know cut a four-hour job down to about seventy minutes. Nothing exotic happened. She opened an AI chatbot online in a browser tab, pasted in a messy client brief, and worked through five deliberate steps instead of typing questions at random. The tool didn’t change. Her method did.
That gap between “I have a chatbot open” and “the chatbot actually saved me time” is where most people stall. They ask one vague question, get a wall of generic text back, shrug, and close the tab. What follows is the workflow I use and teach, with the prompts that matter and the points where it still falls over.
Pick the job before you pick the tool
A chatbot is good at one broad thing: turning messy input into a structured first draft, then reshaping it on request. It is bad at being your memory, your calculator, or your source of truth. Before you open anything, write one sentence on a sticky note. “I want a 200-word summary of this call with action items.” Vague intentions produce vague output, every time.
If you’re still deciding what to open, a rundown of the main AI chatbots online and which tools actually hold up for editing, research and coding is worth ten minutes before you commit to one.
Step 1: Put the role and the limits in your first message
The opening message does more work than any other. It should carry three things: who the chatbot is pretending to be, who the output is for, and what it must not do. Compare these two.
Weak: “Write a product description for a torque wrench.”
Workable: “You’re a technical editor with fifteen years in industrial hardware. Rewrite the paragraph below for factory floor managers who don’t care about marketing language. Under 120 words. Don’t add claims that aren’t in the original.”
The second version takes eight extra seconds to type and removes most of the editing you’d otherwise do. The role sets vocabulary. The audience sets register. The limit stops the model drifting into invented specifications, which is the single most common failure in technical writing.
Step 2: Feed it your real material
Chatbots transform text far better than they generate it from nothing. Instead of asking for a description of a product, paste the actual spec sheet and say “write 90 words for a maintenance engineer, lead with the torque range.” Numbers pulled from real material tend to survive the process; numbers the model produces from memory often don’t.
Same principle for awkward emails. Paste the thread, add one line of context (“this client has already missed two deadlines”), and ask for three reply options ranging from firm to accommodating. You’ll usually get something usable in one pass. Even the strongest models wobble when they have nothing concrete to work from, and where ChatGPT gets things right and where it quietly gets them wrong is a good calibration check on that habit.
Step 3: Ask for options, then make it choose
Never accept draft one. The cheapest upgrade to any output is a small menu. “Give me five subject lines, ranked, with a one-line reason for each.” Then, once you’ve read them: “You ranked number three first. Argue against it and tell me what number five does better.”
That second prompt forces the model into a different mode. It stops producing pleasant-sounding text and starts comparing trade-offs. Nine times out of ten I end up with a hybrid I wouldn’t have written myself.
Step 4: Run a critique pass on its own answer
This is the step most people skip and the one that separates a mediocre result from a finished one. Copy the chatbot’s output back into a fresh message with this: “Find the three weakest claims here and tell me what evidence would be needed to support each one.”
Do not ask “is this good?” It will say yes. Asking it to attack the text produces a genuinely useful list of gaps, usually including the one paragraph you were already suspicious of.
Other second passes worth trying
Ask for a version written at a different reading level, or for a reverse outline (“list the argument of each paragraph in five words”). A reverse outline exposes rambling structure faster than rereading the whole thing.
Step 5: Verify the things that get you in trouble
A chatbot states wrong information in exactly the same tone as right information. There’s no flicker, no hesitation. So build a checklist and run it before anything leaves your desk:
- Dates, prices, statistics, and any legal, medical or financial claim
- Names, job titles and how companies spell themselves
- Citations, links and quotes attributed to a real person
- Numbers that appear twice in the same document, because they frequently disagree
- Code, which should be run rather than read
Assistant-style tools with live search, like Google’s, reduce some of the guesswork but tend to over-compress nuance in a single confident paragraph. A closer look at what Google’s AI chatbot does well and where it lets you down helps you spot when a tidy summary has flattened an important caveat.
A worked example: a 40-minute call to a client-ready summary
Here’s the whole workflow on one task. The input is a 6,800-word transcript of a project kickoff call. The output is a summary email the client can act on. Total time: about 22 minutes.
Minutes 1-4. Paste the transcript in two chunks and ask: “Summarise in eight bullets. After each bullet, quote the sentence that supports it.” The quotes let you skim for accuracy instead of rereading 6,800 words.
Minutes 5-10. “Now list every action item as a table with three columns: owner, task, deadline. Where the owner was never stated, write UNASSIGNED. Don’t guess.” Expect two or three UNASSIGNED rows. Those are the real findings, because they’re the things that will stall the project.
Minutes 11-14. “List every point where two people disagreed, or where a decision was deferred.” This is the part a human summariser usually buries, and it’s the part the client most needs to see.
Minutes 15-22. Write the email yourself from those three outputs. The chatbot did the extraction; you do the judgement and the tone. Skipping that last step is how summaries end up sounding like a committee wrote them.
Mistakes that eat your first hour
- Asking one enormous question instead of a sequence of small ones
- Accepting the first draft’s tone rather than asking for “plainer”, “shorter sentences” or “no adverbs”
- Assuming a long thread is a productive thread. Past fifteen or twenty turns, context gets muddy; start a clean chat and restate what matters
- Never showing an example of good output. “Here are two paragraphs I like. Match this voice” outperforms any adjective you can invent
- Confusing speed with accuracy. The seconds you save skipping verification get paid back with interest
When the workflow becomes something you ship
If the same three prompts keep solving the same problem for your team, that’s a signal you’ve outgrown ad-hoc chatting. Support teams often start by deflecting the repetitive third of their tickets: order status, password resets, return windows. The prompts are easy. The harder work is deciding what the bot is allowed to say, where it hands off to a person, and how you’ll notice when it starts answering confidently and wrongly. That deployment reality is covered well in this breakdown of what works, what fails, and how to deploy an AI bot without regret.
Building for customers is a different exercise again, because a mediocre internal assistant wastes a minute while a mediocre customer assistant wastes trust. If that’s the direction you’re heading, the lessons in building a chatbot that genuinely helps rather than annoys are worth reading before you write a single prompt.
Either way, keep the prompts. Paste the five that work into a shared doc, role and constraints first, critique request second, verification checklist last. The next person on your team starts at step three instead of step zero.