Prompt Engineering

How prompts work, how to write them and how to keep them safe.

GUIDES

Prompt Engineering

Prompt injection: how it works and prompts that harden your assistant

Plain explanation of prompt injection, what OWASP and the UK NCSC advise, and three prompts to copy that make your AI assistant harder to hijack.

 Prompt Engineering

What is prompt engineering? Learn it with prompts you can copy

Learn what prompt engineering is by doing: a weak prompt rewritten, five core techniques with copyable prompts, 2026 best practices and a free way to learn it.

GRAB & GO PROMPTS

Any AI chat

Prompt Improver

Questions about your prompt first, then a rewritten version with a list of what changed.

Act as a prompt editor and improve a prompt I wrote. Do not rewrite it yet. First ask me up to five short questions about anything the prompt leaves unclear: the goal, the audience, the format I want back, the length, and what the AI must not do. Wait for my answers. Then give me the rewritten prompt in a single code block, followed by a short list of what you changed and why. Keep my original intent and keep my wording where it already works. Do not add requirements I did not mention. If something is still unknown, leave a clearly marked blank for me instead of guessing. Quality check: confirm that the rewrite would produce the kind of answer I was originally after, only more reliably. My prompt: [PASTE YOUR PROMPT]
work
Any AI chat

Messy Task to Prompt

Turns a rough description into a prompt with role, context, task, format and constraints.

Act as a prompt designer. I will describe a task in a messy way. Turn it into a structured prompt with these labeled parts: Role, Context, Task, Format and Constraints. If a part cannot be filled from what I wrote, ask me one question about it before you build the prompt, and do not invent the answer. Keep the Task to one clear instruction. Make the Format specific, covering length, structure and tone. Put anything the AI must never do under Constraints. Return the final prompt in one code block that I can paste straight into a chat, then list the assumptions you made in two or three bullet points. Quality check: read the prompt as a stranger would, and fix any sentence that could be understood in two ways. My messy task: [DESCRIBE THE TASK IN YOUR OWN WORDS]
work
Claude

System Prompt Builder

A system prompt for a custom assistant, with scope, tone, limits and sample replies.

Act as an expert at writing system prompts for custom assistants. First ask me about: who will use the assistant, what it should help with, what it must decline or hand off, the tone it should use, and the format its answers should take. Wait for my answers. Then write a system prompt with these sections: Role, Scope, Tone, How to Answer, Never Do, and When Unsure or Out of Scope. Tell the assistant to ask a question when information is missing instead of guessing, and never to invent facts, sources or prices. Add three short sample user messages, each with the reply the assistant should give. Quality check: look for any instruction that contradicts another one, and fix it before you reply. The assistant is for: [PURPOSE AND USERS]
work
Any AI chat

Few-Shot Example Builder

A set of example pairs and a finished prompt, built from your own sample.

Act as a prompt engineer who builds few-shot examples. I will describe a task and give you one or two real examples of the input and the output I want. If I did not give a real example, ask me for one before you continue. Create four more example pairs in the same style: a typical case, an edge case, a very short input and a tricky input. Do not copy my examples. Do not invent facts about my business or data. Use neutral sample content and label it as a sample. Then assemble one complete prompt: a one-line instruction, all the examples labeled Input and Output, and an empty slot for new input. Quality check: confirm that all the examples follow exactly the same format and tone. Task: [WHAT THE AI SHOULD DO] My real example(s): [PASTE INPUT AND OUTPUT]
work
Claude

AI Answer Critic

A point-by-point review of an AI answer against your original request.

Act as a critical reviewer. Below are my original request and the answer an AI gave me. Compare the answer with the request. If it is unclear what I wanted, ask me before you judge. Then report: 1. Which parts of my request the answer covered and which it missed or ignored. 2. Statements that look unsupported or possibly wrong. Mark them 'verify'. Do not call them false unless you are sure. 3. Problems with format, length or tone. 4. What I should add to my prompt to get a better answer next time. Rate completeness, accuracy risk and format fit from 1 to 5, with one sentence of reasoning each. Do not rewrite the answer unless I ask. Quality check: make sure every criticism points to a specific sentence or part of the answer. My request: [PASTE MY ORIGINAL PROMPT] The AI's answer: [PASTE THE ANSWER]
work
Any AI chat

Strict Summarizer

A summary at the length and for the audience you set, with a note of what was left out.

Act as an editor who summarizes a text for one specific reader. If I have not told you the audience, the exact length or the purpose of the summary, ask first and wait for my answers. Then write the summary for that audience at the length I set, and do not go over it. Use only information from the text. Do not add facts, opinions or context from outside it, and keep names, numbers and dates exactly as written. Start with one sentence giving the main point, then the key supporting points. After the summary, add a line starting 'Left out:' that names any important detail you had to drop to fit the length. Quality check: count the words in the summary and state the count. Audience: [AUDIENCE] Length: [EXACT LENGTH] Text: [PASTE THE TEXT]
work
Any AI chat

Ask Me Questions First

The AI asks you questions one at a time and only then starts the task.

I want your help with a task. Do not start it yet. Ask me questions, one at a time, until you have enough information to do it properly. Ask no more than ten. Start with the biggest unknowns: my goal, my audience, my constraints, the format I want and anything I have already tried. Skip anything I have already answered. After each answer, say briefly what you now know, then ask the next question. When you are confident, write 'Ready', give a short summary of what you understood and wait for me to confirm. Never fill a gap with a guess. If I say I do not know, offer two options to choose from. After I confirm, do the task. Quality check: if you reach ten questions and still have gaps, list the gaps instead of guessing. My task: [DESCRIBE THE TASK]
work
Any AI chat

Consistent Output Format

A reusable prompt that makes an AI return JSON or a table in the same shape every time.

Act as a prompt engineer who makes AI answers predictable. I will tell you the task and whether I want JSON or a table. If I have not said which fields or columns I need, ask me first and wait. Then write a reusable prompt that tells the AI exactly what to return: the field or column names, the type and allowed values for each one, what to write when information is missing (null or 'not given', never a guess), and a rule to return only the JSON or the table with no extra commentary. Include one filled example and one example with missing data. Quality check: check your example JSON or table against your own rules and fix any mismatch. Task the prompt is for: [TASK] Format: [JSON OR TABLE] Fields I need: [FIELDS]
work
Any AI chat

Rewrite for Another AI

Your prompt adapted for a different AI model, with each change explained.

Act as a prompt engineer who adapts prompts between AI chat tools. I will paste a prompt that works with one AI model and name the model I want to use it with. If I have not said what the prompt is for, or what went wrong with the new model, ask me first. Rewrite the prompt for the new model. Adjust the structure, the level of detail, the formatting and the wording of the instructions, but keep the goal and all facts unchanged. If you are not sure about a current feature of a model, do not claim it. Write 'check the tool's documentation' instead. Show the new prompt in one code block, then list each change with a one-line reason. Quality check: confirm that nothing from the original instructions was lost. Original prompt: [PASTE THE PROMPT] Original model: [ORIGINAL AI MODEL] New model: [NEW AI MODEL]
work
Any AI chat

Prompt Chain Planner

A big task split into a chain of smaller prompts, each ready to paste.

Act as a workflow designer for AI chat tools. I will describe a big task. If the end goal, the final deliverable or the source material is unclear, ask me first and wait for my answers. Then break the task into 3 to 7 smaller steps, each with its own ready-to-paste prompt. For each step give: the purpose, what to paste in from the previous step, the prompt itself, and the output format to request so that it feeds the next step. Add a short check for me to do between steps wherever a mistake would be costly. Do not invent facts or data inside any prompt. Mark missing inputs as blanks for me to fill. Quality check: confirm that each step uses only what earlier steps produce or what I supply. Big task: [DESCRIBE THE TASK] What I already have: [MATERIALS]
work
Gemini

Fact-Check Helper

A table of the claims in a text and the kind of source that would verify each one.

Act as a careful fact-checking assistant. I will paste a text. List every checkable claim in it (numbers, dates, names, causes, quotes) in a table with these columns: Claim, Type, Risk if wrong, and What would verify it. In the last column, name the kind of source needed, for example official statistics, a primary document or a named expert's published work. Do not say a claim is true unless you are sure. Mark anything else 'unverified'. Never invent sources, links, titles, authors or quotes. If you cannot name a real kind of source, say so. Skip opinions and mark them as opinions. Finish by naming the three claims most worth checking first. Quality check: make sure every row in the table comes from my text and that you added no claims of your own. Text: [PASTE THE TEXT]
learning
Claude

Brand Voice Card

A one-page brand voice card and a short block you can paste into any AI chat.

Act as a brand voice editor. I will give you a brand description and a few writing samples. If I have not told you who the audience is, ask first and wait. Build a one-page voice card with: a three-sentence description of the voice; five tone traits, each with a 'sounds like' and a 'never sounds like' example; preferred and avoided words; sentence length habits; and how the tone shifts between a headline and an email. Base the card only on the samples and the description. Do not invent facts, slogans or values for the brand. End with a short block of instructions I can paste at the start of any AI chat to make it write in this voice. Quality check: write two sample lines in the voice and confirm they match the card. Brand and audience: [BRAND DESCRIPTION AND AUDIENCE] Writing samples: [PASTE 2 TO 3 SAMPLES]
work