Most people use AI chatbots like a search engine that talks back. Ask a thing. Read the answer. Move on. That works. It's also roughly 5% of what these tools are actually useful for.
The people getting real leverage treat them differently. Less like a search box. More like a capable colleague who is fast, tireless, has read enormously widely, and can't be trusted on any specific fact without checking. Sounds like a criticism. It isn't. It's a precise description of what the tool is good for, and knowing that is the whole skill.
So here's the practical version. The seven jobs it genuinely does well. The prompt structure that actually works. What to keep away from it. And how to build the habit so it saves you time instead of becoming another tab you forget about.
The mindset shift that changes everything
The single biggest unlock isn't a prompt. It's realising you are directing, not asking.
A question invites an answer. A brief invites work. Compare:
Write me a LinkedIn post about marketing.
against
I'm writing for founders of Indian D2C brands doing ₹50L–2Cr a year. The point I want to make is that most of them are measuring ROAS at the campaign level and missing that their blended margin has gone negative. Direct, no hype, no emoji, around 150 words. Open with a specific observation, not a question.
The first gets you something generic. Nothing in the request narrows it. The second gets you something usable on the first attempt, because every choice the model would otherwise guess at has already been made for it.
That's the shift. Stop asking, start briefing. Everything below is really just elaboration on that one idea.
The seven jobs it does genuinely well
Notice what these have in common. In every single one, you can judge the output yourself. Not a coincidence. It's the boundary of safe use.
1. First drafts. The blank page problem goes away. A rough draft in twenty seconds that you then rework is faster than staring at nothing. And the reworking is where your voice enters anyway.
2. Summarising material you have. Paste a long thread. A report. A transcript. A set of notes. Ask for the key points. One of the most reliable uses, because everything it needs is right in front of it. It isn't recalling. It's reading.
3. Restructuring. You have the thinking but it's a mess. "Here are my notes, organise them into a logical argument" is genuinely excellent. And much harder to do yourself, because you're too close to it.
4. Explaining something at your level. "Explain attribution windows to me as if I understand marketing but not measurement." Adjustable-depth explanation is a real capability. Often better than the documentation.
5. Being a sparring partner. "Here's my plan. What's the strongest argument against it? What am I assuming?" Underused. And honestly among the most valuable, because it will find holes you can't. You built the thing.
6. Repetitive text work. Reformatting. Rewriting fifty product descriptions in a consistent tone. Converting notes into a table. Cleaning messy data into a usable shape. Dull, high-volume, and the model doesn't get bored.
7. Getting unstuck. Not for the answer. For breaking the paralysis. "Give me ten angles on this, including three bad ones." Even the bad ones move you.
The prompt structure that actually works
Most prompt advice is a list of tricks. This is a structure. It works because it removes the things the model would otherwise have to guess.
1 · Role and context. Who you are, who the output is for, what the situation is.
2 · The task. One clear thing. Not four.
3 · Constraints. Length, format, tone, what to avoid.
4 · An example, if you have one. One sample of "good" beats three paragraphs describing it.
5 · What to do when unsure. "If you need information I haven't given you, ask rather than assume."
That last one matters more than it looks. Without it, a gap in your brief gets filled with something plausible and invented. With it, you get a question instead. Which is what a good colleague would do.
You don't need all five every time. For a quick reformat, the task alone is fine. For anything you'll actually use, all five takes thirty seconds and saves three rounds of correction.
Three habits worth more than any prompt trick
Give it the source material. By a distance the highest-value habit. A model reasoning over a document you supplied is dramatically more reliable than one recalling from training. It removes the entire category of invented facts. If you can paste it, paste it.
Ask for the reasoning before the conclusion. "Show your working, then the recommendation." The reasoning constrains what conclusion is available. And it lets you see where it went wrong, instead of just that it did.
Save what works. When a prompt produces something good, keep it. Within a month you have a small library for the things you do repeatedly. You stop rebuilding the same brief from scratch every time.
A worked example: messy notes to a usable brief
The abstract advice above is easier to see in one real sequence. This is the kind of task where the tool genuinely earns its place. High text volume. Clear success criteria. You can check every part of the output.
The situation. You've come out of a 45-minute call. Two pages of half-sentences. Three action items buried somewhere in them. A brief to write by tomorrow.
Step 1, get it out of your head, unedited. Paste the raw notes in. Don't tidy them first. Tidying is the work you're trying to avoid.
These are my raw notes from a client call. Don't summarise yet. First, list every distinct point you can identify, including ones that seem minor. Mark anything ambiguous with a question rather than guessing what I meant.
The instruction not to summarise yet matters. Ask for a summary immediately and it compresses too early. The thing you half-remembered gets dropped.
Step 2, separate decisions from discussion.
Now split that list into: decisions made, open questions, and actions with an owner. If something doesn't clearly belong in one, put it in open questions.
This is the step you'd do badly yourself. After a long call, everything feels equally important.
Step 3, draft against a shape you specify.
Write this as a one-page brief for someone who wasn't on the call. Context first, then decisions, then actions with owners, then open questions. Plain language, no filler, under 400 words. If a decision is unclear from my notes, flag it rather than smoothing over it.
Step 4, you edit. The output will be roughly right, and wrong in one or two places you'll spot instantly. You were there. It wasn't.
Total: maybe six minutes against forty. And notice where each capability sat. Restructuring in step 2. Drafting in step 3. Your judgement in step 4, which is the only step that couldn't be delegated.
What to keep away from it
Being specific about this is what separates useful from reckless.
Anything where you cannot check the output. This is the general rule. The rest are just examples of it. If you're relying on the tool to know something you don't, you have no way to catch an error.
Facts, figures, citations and dates from memory. It generates plausible-shaped output. A fabricated statistic looks exactly like a real one. There's more on why in the piece on wrong answers. Supply the numbers, or verify every one.
Confidential or client data. Assume anything you paste may be retained and reviewed. Contracts. Personal data. Unreleased financials. Anything under NDA. Do not paste it into a consumer chat interface. Business and enterprise tiers have different data terms. Read them rather than assuming.
Final output that goes out unedited. Not because it reads badly. Because the responsibility is yours and it doesn't know what you know.
Regulated advice. Legal, medical, tax, investment. Useful for orienting yourself and preparing better questions. Not a substitute for someone qualified and accountable.
ChatGPT or Claude?
Honestly, the answer changes every few months. Anyone giving you a firm ranking is describing a snapshot.
More useful than a winner: the differences that persist.
- Both are strong at drafting, reasoning, summarising and code.
- They differ in default tone and in how they handle long documents. Those differences are real enough that people develop a preference for particular tasks.
- Both have paid tiers with better models, larger context and file handling. If you use one daily for work, the subscription is trivially worth it against the time saved.
Pick one and build the habit. The leverage is overwhelmingly in using a tool well and often. Not in having picked the marginally better one. People who switch constantly chasing benchmarks get less done than people who learned one properly.
If you do use both, a reasonable split is this. Use whichever you find better at long-document work for analysis. Use whichever you prefer stylistically for drafting. That's a personal calibration, not a ranking.
Building the habit so it actually sticks
Most people try AI, find it impressive, and then quietly stop. Because they never attached it to a specific recurring task. The fix:
- Pick one task you do every week that involves writing or reading a lot of text.
- Use AI for that one task for a month. Only that one.
- Keep the prompts that worked.
- Then add a second task.
Deliberately slow. It's the approach that survives. Trying to use it for everything at once means evaluating it on tasks it's bad at, concluding it's overrated, and stopping.
A useful signal you're using it well: you're editing rather than accepting. If you take outputs unchanged, you're either doing work too trivial to matter or not checking closely enough. The value is in the first 80% arriving fast, so your judgement goes into the last 20%. Which is where your actual edge lives.
FAQs
Is ChatGPT or Claude better for work?
Both are capable and the ranking shifts every few months. Both handle drafting, summarising, analysis and code well. They differ in tone and long-document handling. Pick one, learn it properly, build a habit. That produces far more value than switching between them chasing benchmarks.
How do I write a good prompt?
Give role and context. One clear task. Explicit constraints on length, format and tone. An example of good if you have one. And an instruction to ask rather than assume when information is missing. Supplying the source material matters more than any phrasing trick.
Can I use AI for client work?
For drafting, structuring and summarising, yes. With your review before anything ships. Don't paste confidential or client-identifying data into a consumer interface. Check the data terms of whichever tier you're on. The output is your responsibility, so the editing step isn't optional.
Will using AI make my writing generic?
Only if you let it produce the final version. Used for the draft and the mechanical work, with your judgement, experience and specifics on top, it does the opposite. It removes the drudgery that was stopping you writing at all.
Do I need the paid version?
If you use it for work more than occasionally, yes. Paid tiers give better models, longer context and file handling. The difference on real tasks is substantial. Against the time saved, it's one of the cheaper tools you'll pay for.
What should I never use AI for?
Anything you can't verify yourself. Facts, figures and citations recalled from memory rather than supplied by you. Confidential or client data in a consumer interface. Regulated advice. And anything shipping to a client unedited.
Key takeaways
- Direct it, don't ask it. Brief in, work out. A question gets you generic output.
- Seven reliable jobs: drafting, summarising, restructuring, explaining, sparring, repetitive text work, getting unstuck. All share one property, you can judge the result.
- Structure the prompt: role, task, constraints, example, and what to do when unsure.
- Supplying the source material is worth more than every prompt trick combined.
- Never use it where you can't check the output.
- Pick one tool, attach it to one recurring task, expand from there.
Related reading: how to automate your work with AI, why ChatGPT gives wrong answers, and the skills that actually matter in the age of AI.
Top comments (0)