0 / 7
System Prompt
The background instruction you give an AI before any conversation starts. It sets the AI's role, knowledge, and behaviour for all your interactions — like a job brief that's always active.
Write a system prompt that defines your role and business
Tell the AI who you are, your industry, your goals, and the types of tasks you typically do. This is your "background briefing" that persists across every conversation.
HOW +
- Open your LLM's settings — in ChatGPT: Settings → Personalization → Custom Instructions. In Claude: create a Project and set the system prompt there.
- Write 4–6 sentences covering: your role and company, your industry, the tasks you most commonly do, your preferred tone, and any key terminology.
- Save and test by starting a new conversation — the AI should already "know" you without you explaining yourself.
"I'm a marketing manager at [Company], a B2B SaaS company in the HR tech space. I primarily work on content, paid campaigns, and marketing analytics. Always use a clear, professional but conversational tone. Avoid jargon."
Memory
Short-term: context the AI remembers within one conversation — resets when you start a new chat. Long-term: persistent information kept between sessions — your preferences, past context, recurring instructions.
Activate short-term and long-term memory
Enable memory so the AI learns from your conversations over time. Short-term helps within a session; long-term builds a profile of your preferences and context.
HOW +
- ChatGPT: Settings → Personalization → Memory → toggle on. The AI will start saving facts about you automatically.
- Claude: Memory is built into Projects — add key facts manually in the project instructions.
- Periodically review saved memories (ChatGPT: Settings → Personalization → Manage Memory) and delete anything outdated or incorrect.
- Prompt the AI to remember something explicitly: "Remember that I always want outputs in British English and under 200 words."
Retrieval / RAG
Retrieval-Augmented Generation — the AI searches through your uploaded documents to find relevant facts before answering, rather than relying on training data alone. Eliminates hallucination for known information.
Upload key documents and data the AI can retrieve
Add files like pricing sheets, feature lists, campaign timelines, brand guidelines, or FAQs. This is your AI's reference library — it retrieves what it needs rather than guessing.
HOW +
- Gather your core reference documents: brand guidelines, pricing sheets, product features, campaign briefs, tone of voice guide.
- ChatGPT: create a Project and upload files there. Claude: upload to a Project. Files stay attached for every conversation in that project.
- Test it: ask the AI a specific question only answerable from your documents (e.g. "What's our pricing for the Pro tier?"). If it answers correctly, retrieval is working.
- Update files whenever information changes — stale documents cause confidently wrong answers.
Tools & Connectors
Integrations that give the AI real-time access to external apps and data — Google Drive, calendar, CRM, meeting recorders, and more. The AI can read and sometimes act on these sources directly.
Connect relevant tools — Google Drive, calendar, meeting recorder, CRM
Use native connectors in ChatGPT, Gemini, or Claude to give the AI live access to your workspace. The more context it can pull automatically, the less you have to type.
HOW +
- ChatGPT: Settings → Connected Apps → connect Google Drive, OneDrive, or others.
- Claude: use Integrations inside a Project to connect services.
- Start with tools you use daily: Google Drive for documents, Google Calendar for scheduling context, your CRM if available.
- Test by asking the AI to retrieve a specific file: "Find the Q1 campaign brief in my Google Drive."
Output Formats
The structure, length, and style you want the AI to produce — e.g. a bullet-point brief, a specific document template, a JSON object. Without explicit examples, the AI defaults to a generic internet-average format.
Define your preferred output formats and add examples
If you want a brief in a specific structure, a post in a certain style, or a report with particular sections — show the AI exactly that. Without examples, output will be generic.
HOW +
- Find 2–3 real examples of outputs you love — a brief, a post, a report — from your own work or a benchmark you admire.
- Add them to your system prompt under a section called "Output examples:" or upload them as reference files.
- For recurring formats, create a named template: "When I ask for a campaign brief, always use: Overview, Audience, Message, Channels, KPIs, Timeline."
- For one-off tasks, paste the example directly in your prompt: "Format the output exactly like this: [paste example]."
Rules & Guardrails
Explicit instructions that constrain the AI's behaviour — what it must always do, never do, and when to pause for human review. This is how you maintain brand standards and quality control at scale.
Set rules: what's allowed, what's not, and when a human must review
Specify tone rules, off-limits topics, brand restrictions, and at what points a human needs to approve before something goes out.
HOW +
- List your non-negotiables: always use British English, never claim statistics without a source, always include a CTA, etc.
- List what should trigger a human review: anything going to external audiences, legal or compliance topics, pricing decisions.
- Add these to your system prompt under "Rules:" — keep them numbered and specific.
"Rule 3: Never mark any copy as publish-ready. Always end responses that need sign-off with [REVIEW NEEDED]."
Model Selection
Choosing which AI model (e.g. GPT-4o, Claude Sonnet, Gemini Flash) to use for a given task. Models vary in reasoning depth, speed, cost, and specialisation. Using the wrong model wastes money or produces weak results.
Choose the right model for each type of task
Not every task needs the most powerful model. Match the model to the task — complex reasoning needs a different model than quick summaries or image generation.
HOW +
- Complex reasoning, long documents, nuanced writing → GPT-4o, Claude Sonnet/Opus, Gemini Pro.
- Quick tasks, summaries, simple rewrites → GPT-4o-mini, Claude Haiku, Gemini Flash. Faster and cheaper.
- Image generation → DALL·E 3, Midjourney, Ideogram, Flux.
- Video generation → Sora, Runway, Kling.
- When using agentic tools (Manus, Lovable) — check settings for model selection to control cost per task.
0 / 8
Core Habit
Write prompts like you're briefing a capable teammate, not issuing a command
Instead of "write a landing page", give full context: task, background, format, and constraints. Full detail upfront prevents expensive back-and-forth.
HOW +
- Before typing, mentally answer: What do I want? Why do I need it? What constraints matter? What should the output look like?
- Use this template: [Task]. Context: [background]. Format: [how you want it]. Constraints: [word count, tone, what to avoid].
- Read your prompt back before sending — if a human intern wouldn't understand it, the AI won't either.
"Write a LinkedIn post about our new product launch. Context: We're launching a CRM integration that saves sales teams 3 hours/week. Format: 150 words, conversational, end with a question to drive comments. Avoid buzzwords like 'game-changer'."
Input Method
Use voice input for richer, more personal prompts
Speaking is faster than typing and naturally produces more detailed, expressive prompts — especially useful for content you want in your own voice.
HOW +
- Mac: press Fn Fn to activate built-in dictation, or install Whisper for higher accuracy.
- ChatGPT desktop/mobile: tap the microphone icon in the input field.
- Claude: use your system's dictation tool, then paste into Claude.
- Speak naturally — include context, reasoning, and nuance you'd normally skip when typing. The extra detail is the point.
Input Method
Use screenshots and images to show rather than tell
Visuals communicate layout, style, and intent faster than text. Use competitor ads, diagrams, website screenshots, or photos of hand-drawn sketches.
HOW +
- Mac: Cmd + Shift + 4 to capture a selection. Cmd + Shift + 5 for more options.
- Windows: Win + Shift + S.
- Drag the screenshot directly into the chat input, or use the attachment icon.
- Use for: competitor ads, layouts to replicate, error messages, wireframes, handwritten notes, or any visual reference for style.
Structured Prompting
A formal prompt format with defined sections: Role, Instruction, Context, Output Format, and Edge Cases. Critical for complex tasks and agentic tools where back-and-forth is expensive.
Use the academic prompt format for complex tasks and agentic tools
Structure: Role → Instruction → Context → Output format → Edge cases. Essential in tools like Lovable and Manus where iterations cost credits.
HOW +
- Use this template every time you work with an agentic tool or a complex deliverable:
"You are a [role] with [X years of experience in Y].
Your task is to [specific instruction].
Context: [relevant background].
Output format: [exactly how you want the result].
Constraints: [word count, tone, things to avoid].
Edge cases: [exceptions to handle]."
Your task is to [specific instruction].
Context: [relevant background].
Output format: [exactly how you want the result].
Constraints: [word count, tone, things to avoid].
Edge cases: [exceptions to handle]."
- Save your best versions as reusable templates — one per recurring task type.
- The more expensive the tool in credits or time, the more detailed your prompt should be.
Meta-Prompting
Using one AI to write or improve a prompt that you'll then use with a different AI or specialised tool. Bridges the gap between your intent and the ideal input format for the target model.
Use one LLM to write or refine prompts for another
Especially useful for image/video generation or specialised tools where prompt syntax is highly specific and you're not an expert in the domain.
HOW +
- Describe your idea or goal to your main LLM (Claude or ChatGPT) in plain language.
- Ask it to refine and clarify the concept with you.
- Say: "Now write a detailed prompt I can use in [Midjourney / Sora / Manus / Lovable] to achieve this."
- Review the generated prompt, adjust if needed, then paste into the target tool.
Quality Control
Use a second LLM to critique and improve output from the first
Don't stay in one chatbot that keeps praising everything. Cross-model critique consistently raises output quality.
HOW +
- Complete your task in your primary LLM and copy the output.
- Open a different LLM (if you used ChatGPT, try Claude or Gemini).
- Paste the output and ask: "What are the 3 weakest parts of this? How would you rewrite them?"
- Or ask: "What would make this 20% better?"
- Combine the best of both outputs manually.
Reverse Prompting
Asking the AI to explain what prompt would have produced a better output. Useful for diagnosing why output missed the mark, or for reverse-engineering a style or format you admire.
When output is wrong, ask the AI: "How should I have prompted you?"
Diagnose before you regenerate. AI is surprisingly good at explaining the gap between your intent and what it received.
HOW +
- When output is wrong, say: "This isn't right. I was hoping for [describe what you wanted]. Why didn't my original prompt produce that, and how should I have phrased it?"
- Use it proactively: show the AI an output you admire and ask "What prompt would produce something like this?"
- Save the improved prompt for next time — that's your new template.
Dialogue
Ask the AI to ask you clarifying questions before it starts
Especially important with ChatGPT, which tends to proceed with assumptions. Surfaces ambiguity early and avoids wasted outputs.
HOW +
- Add this line to any prompt where you're unsure if you've given enough context: "Before you begin, ask me up to 3 clarifying questions that would help you do this better."
- Answer the questions, then let the AI proceed.
- For ChatGPT: add this as a rule in your system prompt — "Always ask clarifying questions before starting any task that involves creating new content."
- Especially valuable for complex deliverables: landing pages, strategy docs, research reports.
0 / 9
I have a system prompt set up
Memory is activated
I've uploaded relevant files
Tools & connectors enabled
Output format examples ready
Rules & guardrails defined
I write detailed user prompts
Human review points specified
I use a second LLM to critique