On this page
- The seven techniques, ranked by how often you'll use them
- 1. Goal, audience and constraints
- 2. Supply the source material
- 3. Few-shot examples
- 4. Control the output format
- 5. Reasoning prompts
- 6. Self-critique and revise
- 7. Prompt chaining
- Techniques to be careful with
- Differences between Claude, ChatGPT and Gemini
- A workflow for turning techniques into habits
- Next step
The prompting techniques that pay off most are: state the goal and audience, give the model the source material, show one or two examples, specify the output format, and run a self-check pass. Fancier methods like chain-of-thought and prompt chaining help with specific problems, and you can skip them for the rest.
This post covers seven techniques in the order we'd reach for them. Each has a prompt you can paste. Where the advice depends on the model, we say so.
The seven techniques, ranked by how often you'll use them
Goal, audience and constraints
Supplying source material
Few-shot examples
Output format control
Reasoning prompts (step by step)
Self-critique and revise
Prompt chaining
The first four fix most weak outputs. Techniques five to seven are for harder tasks: analysis, long documents, multi-part deliverables.
1. Goal, audience and constraints
Tell the model what the output is for, who reads it, and what limits apply. This is the cheapest improvement you can make, and most people skip half of it.
Here's a weak prompt:
Write an email to a customer about a delayed shipment.
The answer will be polite, generic, and padded with apologies. Now the same task with a goal, audience, and constraints:
Write an email to [customer name], who ordered [product] on [date]. The shipment is delayed by [number] days because of [reason]. Goal: keep the customer from cancelling by being honest and giving them a clear choice. Audience: a busy buyer who is mildly annoyed, not furious. Constraints: under 120 words, no more than one apology, offer two options: wait for the new date of [new date] or get a refund. Plain tone, no corporate phrases.
The second version gets a short email that leads with the news, names the new date, and ends on a choice. The changes that mattered: the word limit, the single apology and the two explicit options. If you want to go further on this approach, Prompting in 2026: Write Goals, Not Rulebooks explains why a clear goal beats a long list of rules.
2. Supply the source material
Models invent details when you leave gaps. If the answer depends on facts, paste the facts in and tell the model to stay inside them.
Use only the text between the <source> tags to answer. If the answer isn't in the source, say "Not in the source" instead of guessing. <source> [paste document, notes or data] </source> Question: [your question]
The delimiter tags aren't magic. They just stop the model from confusing your instructions with your material. Claude tends to handle XML-style tags well, and ChatGPT and Gemini cope fine with them too, so one habit works across all three.
The "say so if it's not there" line is the part doing the work. Without it, a model will often fill the gap with something plausible.
3. Few-shot examples
Showing one to three examples of the output you want is often faster than describing it. It works best for tone, tagging, classification, and anything with a house style.
Rewrite customer feedback into a one-line bug report in this style.
Example input: "The app keeps crashing when I try to upload a photo from my gallery on my phone, super annoying"
Example output: "Crash on photo upload from gallery (mobile)."
Example input: "I can't find where to change my billing email, went through settings twice"
Example output: "Billing email setting hard to locate."
Now rewrite this:
[feedback text]Two things to watch. Models copy your examples closely, so if both are short, the output will be short. And if your examples share an accidental quirk, such as both starting with a noun, the model will pick that up as a rule. Vary them on purpose. We go into the mechanics in Few-Shot Prompting: What It Is and How to Use It.
4. Control the output format
Name the structure you want: a bulleted list, three headings, JSON with specific keys, a two-column list written as plain lines. Formatting instructions are cheap, and models follow them reliably.
Summarise the meeting notes below.
Format:
- Decisions: bullet list, one line each
- Open questions: bullet list, with an owner in brackets if one was named
- Next steps: numbered list with a due date where given
If a section has nothing, write "None". Don't add any other sections.
Notes:
[paste notes]The "don't add any other sections" line matters. Left alone, models add a friendly intro and a closing summary you didn't ask for.
For developers who need machine-readable output, check what your model provider offers for enforcing a schema, since the options differ by provider and change often. Asking for JSON in the prompt alone works most of the time, not all of the time.
5. Reasoning prompts
Asking a model to work through a problem step by step can improve answers on maths, logic, planning and multi-factor decisions. It's less useful for simple rewriting or lookup tasks, where it only adds length.
Some newer models do their own reasoning before replying, and for those a "think step by step" instruction may add little or nothing. Check how your model behaves. Test the same prompt with and without the line and compare.
Here's a version that asks for reasoning you can actually check:
I need to decide [decision]. Options: [option A], [option B], [option C]. What matters most: [criteria, in priority order]. First, list the facts you're relying on and flag any assumption you're making. Then compare the options against each criterion. Finish with a recommendation in two sentences.
The useful part is the assumptions list. It shows you where the model made something up, which is where you correct it.
6. Self-critique and revise
Get a draft, then ask the model to review it against specific criteria and rewrite. Vague requests like "make it better" produce cosmetic changes. Specific criteria produce real edits.
Here's a draft [type of document]: [paste draft] Review it against these criteria: 1. Does the first sentence state the main point? 2. Is every claim backed by something in the draft? 3. Is there any sentence that repeats the previous one? List the problems you find, quoting the exact text. Then give a revised version that fixes only those problems and leaves everything else untouched.
"Leaves everything else untouched" stops the model from rewriting your voice out of the piece. For writing tasks, ChatGPT for Writing: A Practical Workflow With Prompts shows this critique step inside a full draft-to-edit workflow.
7. Prompt chaining
Split a big task into a sequence of smaller prompts, each one feeding the next. Chaining is worth it when one prompt is trying to do too much, like researching, structuring, and writing a report in a single go.
A typical chain for a client proposal:
Extract the client's stated needs from the brief into a list.
Turn that list into an outline, with one section per need.
Draft each section separately.
Run the critique prompt from technique 6 over the full draft.
You can review the output at every step and fix a bad outline before it becomes a bad draft. The cost is time and more copy-pasting. If the task fits one clear prompt, don't chain it.
Techniques to be careful with
Role prompts alone. "You are a world-class marketer" adds little by itself. A role helps when it comes with the audience and the goal, as in technique 1.
Emotional pressure. Lines like "this is very important for my career" are folklore. Don't count on them.
Very long rule lists. Past a dozen rules, models start dropping some. Cut the list down to the rules that matter.
Copying a technique because it's famous. If a plain prompt already gives you what you need, adding a technique is just more words.
Differences between Claude, ChatGPT and Gemini
The core techniques carry across models. The differences are at the edges.
Claude responds well to clearly tagged sections and explicit instructions about what to do when information is missing. For Claude-specific templates, see Prompts for Claude: 7 Templates You Can Paste Today.
ChatGPT is quick to add intros and wrap-up lines, so format constraints and "no other sections" instructions earn their place.
Gemini handles long pasted material and mixed inputs well, though you should still say exactly what to do with each part.
These are tendencies, not guarantees, and they shift as models are updated. The only reliable method is to run your prompt on the model you'll use. If you work across several, How to Get Consistent AI Results Across Claude, ChatGPT, and Gemini covers how to keep outputs comparable.
A workflow for turning techniques into habits
Knowing the techniques is one thing. Applying them every time is another, mostly because good prompts end up scattered across chats and documents.
Pick a task you do weekly, such as a status update or a first-pass email reply.
Write the prompt using techniques 1 to 4: goal, source, example, format.
Run it five times on real inputs and note where it fails.
Add one fix per failure, such as a constraint, a second example, or a "say so if missing" line.
Save the final version with placeholders for the parts that change.
PromptPal's prompt library is where we keep those finished prompts, and prompt version control lets you see which edit made the output better or worse. That matters more than it sounds: after three tweaks, nobody remembers which wording worked.
Next step
Pick one prompt you use every week and rewrite it with the goal, source, example, and format from this post. Then save it in the public prompt library as a starting point, or browse existing templates by role and task to see how others have structured theirs.
Stay Updated
Get the latest blog posts and AI prompt engineering tips delivered to your inbox.




