How to Write Better Prompts for Claude and ChatGPT
Most disappointing AI answers trace back to one thing. Vague prompts. Not weak models. I’ve watched people blame the tool when the real problem was the ask. So here’s a repeatable method I lean on — the one that lands a useful answer on the first try, most of the time.
📌 Related reading: Digital Nomad Lab · Home Tech Lab
In This Guide
Steps
Step 1: State the Goal, Not Just the Task
Don’t just say “write an email.” Say “write a short email asking my landlord to fix a leaking faucet, polite but firm.” See the difference? That goal, plus that tone, changes the output dramatically. Hand a model only the task and it guesses at your tone and your audience. It usually guesses generic. The more you pin down up front, the less editing you’ll grind through later.
Step 2: Give the Model Relevant Context Up Front
Paste in the actual document, the code, the background you’re working with. Don’t describe it from memory. Models do far better with the source material sitting right there in the prompt. I learned this the hard way. I once summarized a long report from memory, then did it again by pasting the real text — and the pasted version kept catching details I’d otherwise have left out or misremembered entirely.
Step 3: Specify the Format You Want Back
Want a table? A bullet list? A specific word count? Then say so. Plainly. Models default to prose unless you tell them otherwise, and prose often isn’t what you actually need. One small line does most of the work here — “return this as a markdown table with three columns” — and almost all the ambiguity about shape just falls away.
Step 4: Ask for One Task at a Time on Complex Requests
Multi-step work? Split it. Into a sequence of prompts. One giant request buries your errors until the very end, where they’re hardest to catch. A single sprawling prompt that demands research, an outline, and a final draft all in one breath tends to hand back a weaker version of each. Three focused prompts, run in sequence, beat it almost every time.
Step 5: Iterate Instead of Starting Over
First response lands close but not quite? Tell the model exactly what to change. Don’t rewrite your original prompt from scratch. Rewriting tosses out the useful context you built in that first attempt. Something short usually does it — “keep the structure, but make the tone more casual” — and out comes a workable second draft.
Step 6: Build a Small Library of Prompts That Already Work
Hit a prompt structure that reliably works for some recurring task — a weekly summary, a code review request, a travel-planning brief? Save it. As a template. Don’t rebuild it from memory every single time. This pays off across any repeating workflow, whether you’re drafting emails at a desk job or digging into visa and cost-of-living details as a digital nomad comparing cities on a budget.
Tips
- ✅ Save prompt templates that work well for recurring tasks (emails, summaries, code reviews).
- ✅ Need to verify a factual claim? Ask the model to explain its reasoning briefly first.
- ✅ Lean on explicit constraints — “under 150 words,” “no bullet points” — whenever format actually matters.
- ✅ The same specificity rule covers home automation prompts too, like naming routines or writing scene descriptions for a smart speaker. A vague “make my evening routine” gets a far weaker result than naming the exact devices and the exact times. That precision helps if you’re also trying to reduce your smart home’s standby power usage through smarter scheduling.
Warnings
⚠️ Always verify factual claims, statistics, or citations an AI model hands you. Models can produce plausible-sounding but flat-out wrong information. ⚠️ Be careful about pasting sensitive personal, financial, or client data into a prompt unless you’re confident about the platform’s data handling policy.
Q&A
Does prompt length matter?
Not on its own, no. A longer prompt packed with relevant context usually beats a short vague one. But pad it with irrelevant detail and you gain nothing.
Is there a difference in prompting Claude vs. ChatGPT?
The core principles carry across both. Clear goal. Real context. Defined format. Each model has slightly different strengths, though, so it’s worth testing them against your specific use case.
Should I use the same prompt style for coding tasks and writing tasks?
Mostly, yes — the core structure holds either way. Coding prompts just gain even more from pasting the exact error message, or the real code, instead of paraphrasing it.
Frequently Asked Questions
What if the model ignores part of my prompt?
That usually means too many competing instructions are crammed into one request. Split the task into steps or number your requirements, and put the single most important constraint last. Both Claude and ChatGPT weight the most recent instructions heavily, so ordering matters.
Does adding examples to a prompt always help?
Not always. One or two clear examples usually improve format-following, but piling on more can make the model copy them too literally instead of reasoning. How much example context helps varies by model, so test it on your own task rather than assuming.
What if I get a different answer every time I run the same prompt?
Some variation is expected because these models sample their responses. Where a tool exposes a “temperature” setting, lowering it makes output more consistent; for near-identical repeats, set it to the lowest value the platform allows. Exact reproducibility is not guaranteed on most consumer chat interfaces.
Sources: Anthropic — Claude Prompt Engineering Overview, OpenAI — Prompt Engineering Guide. Fact-checked based on public sources as of July 20, 2026.
