Prompting is not magic words. It is giving a very capable, very literal colleague the context they need. Here is how it works, what actually changes the output, and 47 prompts you can copy today.
18
Techniques
14
Influence factors
8
Guides distilled
47
Library prompts
Prompt upgrade
Before
write a blog post about remote work
After
Learn these and most prompting advice becomes obvious. Each one is distilled from how the labs themselves describe their models.
The model only knows what is in front of it. Everything it sees shapes the answer, not just your last message.
A language model predicts a response from everything in its context window: the system prompt, any documents, tool results, earlier turns and your latest request. If a fact, preference or constraint is not in that window, the model does not know it.
That is why the same question gets better answers in a well-set-up project than in a blank chat. Good prompting is less about magic words and more about giving the model the right material to work with.
Each coloured part does a job. Click one to see what it contributes, or switch it off to watch the predicted output fall apart.
The prompt
Perspective
Sets voice and point of view. Useful, but the weakest part of this prompt on its own.
Predicted output quality
100/100
What you tend to get
Subject options: 1) Lost signal? Keep working. 2) Offline Mode is here 3) The basement problem, solved Starting today, you can keep working when the signal drops. Jobs, photos and signatures save on your device and sync the moment you're back online, with no lost work and no re-entry. You asked for this more than anything else, and we took the time to get it right… Preview: Jobs and photos now save offline and sync automatically.
Not all prompt elements are equal. On 2026 frontier models, context and examples dwarf personas, shouting and magic phrases. Open any row for a weak and strong example.
Tells the model what to optimise for. Without it, the model optimises for a generic, safe answer.
Weak
Write something about our new feature.
Strong
Write a 120-word changelog entry announcing offline mode. Success: a busy user understands what changed and why they should care in one read.
Tip: End your brief with one sentence describing what a great result looks like.
Influential phrases
Short phrases that reliably change behaviour. Swap the braces for your details.
Honesty
“If you're unsure, say so instead of guessing.”
Cuts confident hallucinations and invites calibrated answers.
“Quote the relevant passage before answering.”
Grounds answers in your documents and makes them checkable.
Scope
“Only use the information provided.”
Keeps the model from mixing in outside or outdated knowledge.
“Don't change anything outside {{scope}}.”
Prevents helpful but unwanted extra edits.
“Ask me up to three questions before you start.”
Surfaces missing context on ambiguous tasks.
Quality
“Before answering, list the requirements and check each one.”
Self-verification catches missed instructions.
“Give me two options and recommend one.”
Turns vague advice into a decision with a rationale.
“Write it for {{audience}} who will {{action}}.”
Calibrates depth, tone and vocabulary automatically.
Structure
“Return only the {{artifact}}, with no preamble.”
Clean, paste-ready output.
“Answer first, then explain.”
Front-loads the useful part for busy readers.
“Use this exact format: …”
Stable structure across many runs.
Reasoning
“Consider at least two approaches before choosing.”
Reduces anchoring on the first idea.
“What would make this answer wrong?”
Prompts a critique pass that exposes weak assumptions.
“Work through it carefully; accuracy matters more than speed.”
Encourages deeper reasoning on hard problems.
Filter by level. Impact bars show how much each technique typically improves results when it fits the task.
Ask directly, with the goal, audience and format spelled out. Most tasks need nothing more.
Use when: Everyday tasks where the model already knows how to do the work.
Show two to five input and output pairs so the model learns the pattern.
Use when: Classification, extraction, house style and any task with a specific format.
⚠ Models copy examples closely. Vary them or the output will be monotonous.
Paste the documents, background and constraints first, then ask the question at the end.
Use when: Questions about long documents, research synthesis, contract review.
Specify the exact output shape: a table, headings, or JSON that matches a schema.
Use when: Anything a program will parse, or reports that must look the same every time.
Separate instructions, context, examples and data with named tags.
Use when: Long or reusable prompts, and whenever pasted content could be mistaken for instructions.
Invite the model to ask what it needs before it starts.
Use when: Ambiguous or high-stakes tasks where you are not sure what context matters.
Break a big task into steps, each with its own prompt, feeding outputs forward.
Use when: Complex deliverables where single prompts drop details.
Ask the model to review its own draft against criteria, then improve it.
Use when: Writing, code and analysis where quality matters more than speed.
On reasoning models, raise the effort setting and describe the goal instead of prescribing steps.
Use when: Math, debugging, planning and analysis on Opus 5.5, GPT-6, Gemini 4 and similar models.
⚠ Detailed step-by-step scripts can make reasoning models worse.
Make the model extract supporting quotes before it answers.
Use when: Q&A over documents where accuracy and traceability matter.
Ask a strong model to write or improve the prompt for you.
Use when: Starting a new prompt, or improving one that underperforms.
Run the same prompt several times and take the majority or best answer.
Use when: High-stakes reasoning where an occasional error is costly.
⚠ Multiplies cost and latency.
Use a model as a grader with an explicit, specific rubric.
Use when: Automated evals, content QA and comparing prompt versions.
Show a bad example next to a good one and explain the difference.
Use when: Teaching style and judgement that is hard to describe.
Start the model's answer for it to lock in format or skip preamble.
Use when: Strict formats in API use, when structured output modes are not available.
⚠ Not supported in every API or with every reasoning mode.
For agents and multi-hour tasks, write a brief with goals, state files and verification rules.
Use when: Coding agents, research agents and any run longer than a few minutes.
Describe subject, setting, composition, lighting, style and any text, in plain sentences.
Use when: Any text-to-image model; modern ones prefer natural sentences over keyword lists.
Write like a director: shot type, camera motion, action beats, lighting and sound.
Use when: Veo, Sora, Kling, Runway and other text-to-video models.
The fundamentals transfer, but each model has quirks. Here is the single most useful tip for each leader; the full profiles have ideal prompts to copy.
We read the official guides so you don't have to. One standout idea from each, and the full digest a click away.
Prompt library