Master the Prompt — Learn how ChatGPT interprets prompts and why structure matters.
The definitive guide · 2025/2026 edition
How to write the perfect ChatGPT prompt
A technically accurate, beginner-friendly guide to how ChatGPT actually reads and processes what you type — and how to use that knowledge to get dramatically better answers.
Most people treat ChatGPT like a search engine with better grammar. They type a vague question, get a vague answer, and conclude "AI isn't that useful." This guide fixes that. By the end, you'll understand how tokens work, what happens inside the context window, why some prompts fail, and how to write prompts that get consistently excellent results.
Part 1
What ChatGPT actually is
What is a large language model?
ChatGPT is built on a large language model (LLM) — specifically, OpenAI's GPT series. An LLM is a neural network trained on enormous amounts of text. During training, it learned to predict: given a sequence of text, what word (or word-piece) comes next? Do that billions of times across billions of examples, and the model develops a compressed statistical understanding of language, facts, logic, and patterns.
ChatGPT doesn't "think" or "understand" the way a human does. It generates the most statistically likely continuation of your text, given everything it learned in training and everything you've told it so far. That's not a weakness — it's an extremely powerful capability when you know how to use it.
What happens after you press Enter?
When you send a message, here is a simplified version of what occurs:
The word "transformer" in "GPT" stands for the architecture underlying ChatGPT. Transformers process tokens not one at a time (like older models did) but all at once — each token can "pay attention" to every other token in the context. This is what gives ChatGPT its ability to understand context, follow complex instructions, and maintain coherence across long conversations.
What is a token?
ChatGPT doesn't read your text character by character or word by word. It reads tokens, which are chunks of text that the model learned to treat as single units. Common short words are usually one token. Longer or rarer words may be two or three tokens. As a rough guide, one token ≈ ¾ of an English word, so 1,000 tokens ≈ 750 words.
Why does this matter? Because every limit in ChatGPT — the context window, pricing, and processing time — is measured in tokens, not words or characters.
Does ChatGPT read top to bottom? Does it read your whole prompt first?
Yes and no. Before generating any output, the model processes your entire prompt (plus all prior conversation history) via the attention mechanism. Every token in the input can "attend to" every other token simultaneously. So in one sense, it reads everything at once. However, it then generates the response token by token, left to right. Each newly generated token is influenced by the full input plus all already-generated output tokens.
Practically speaking: the beginning and end of your prompt tend to get more attention. Research has shown that information buried in the middle of very long prompts gets lower attention weights, a phenomenon sometimes called the "lost in the middle" problem. This is why critical instructions should appear at the start or end of your prompt, not buried in the middle.
The context window
The context window is the total amount of text (measured in tokens) that ChatGPT can process at one time. It includes everything: system instructions, your entire conversation history, uploaded files, and the current message.
| Model | Approximate context window | Rough equivalent |
|---|---|---|
| GPT-4o (ChatGPT default, 2024–2025) | 128,000 tokens | ~96,000 words / ~300 pages |
| GPT-5 (current, 2025) | Up to 400,000 tokens input | ~300,000 words / ~900 pages |
| ChatGPT Plus/Pro "Instant" mode | 32,000–128,000 tokens (plan-dependent) | ~24,000–96,000 words |
| ChatGPT Plus/Pro "Thinking" mode | 196,000 tokens | ~147,000 words |
These numbers are from publicly available documentation and may change. The model that appears in ChatGPT's interface may differ from the raw API model. Always check OpenAI's current documentation for exact limits. Token limits also apply jointly to input AND output — a very long response eats into the space available for your conversation history.
Why very long prompts become less effective
Several factors degrade quality as prompts grow longer:
- Lost in the middle: Attention weights thin out for content in the centre of a very long context. Crucial instructions may get underweighted.
- Noise dilutes signal: More text means more opportunities for irrelevant content to distract the model from what you actually need.
- Truncation: If the conversation exceeds the context window, the oldest content gets cut. ChatGPT may "forget" things you said early in a very long chat.
Conversation memory vs saved memory
There are two distinct systems:
- Conversation context: The rolling history of the current chat. Everything you and ChatGPT have said in this session is re-sent to the model with each new message. This is why ChatGPT can refer back to what you said earlier — it literally re-reads the whole conversation every time.
- Saved memory ("Dreaming"): A separate system, rolling out to Plus/Pro users in 2025–2026, where ChatGPT synthesises useful facts from your past conversations and stores them externally. This memory is injected into the system prompt at the start of each new conversation. You can view, edit, and delete these memories in ChatGPT's settings.
ChatGPT does not have a persistent brain that accumulates knowledge about you automatically and indefinitely. What it "remembers" is either within the current context window, or what has been explicitly stored in the memory system. If memory is off, every new conversation starts completely fresh.
Why attachments aren't "magically understood"
When you upload a PDF, image, or file, it isn't processed by some separate magical AI — it's converted into tokens and added to the same context window as everything else. A PDF's text is extracted and included as text tokens. An image is encoded as a series of visual tokens. These compete for space in the same context window as your conversation. Understanding this explains why very large files can hurt response quality: they consume context space that would otherwise hold your conversation history and instructions.
Part 2
How ChatGPT reads your prompt
When you send a message to ChatGPT, the model doesn't receive just your typed text. It receives a structured package of context. Conceptually, the full input looks like this:
Why earlier messages affect later answers
Because the entire conversation history is sent to the model with every new message. If you told ChatGPT in message 3 that "this is for a 5-year-old audience," that instruction is still in the context when you send message 15. The model will still (try to) follow it.
Why conflicting instructions create worse results
Imagine you say "write formally" in message 1, then "be casual and funny" in message 10. Both instructions are in the context. The model attempts to satisfy all constraints simultaneously, which often produces incoherent, fence-sitting output. Clear, non-contradictory instructions are one of the most important things you can do for prompt quality.
If you want to override a previous instruction, say so explicitly. "Ignore my earlier instruction about tone. From now on, write in a casual voice." This removes ambiguity about which instruction should win.
What the model actually weighs
The attention mechanism doesn't give equal weight to all tokens. Generally:
- The most recent user message gets strong attention.
- The system prompt (OpenAI's instructions) has strong influence because it was specifically trained to be followed.
- Content near the start and end of long contexts gets more attention than content buried in the middle.
- Repeated instructions across multiple turns carry cumulative weight.
Part 3
The anatomy of the perfect prompt
Every high-quality prompt contains some combination of these components. Not every prompt needs all of them — but knowing each one lets you decide what to include.
1. Role
Telling ChatGPT to adopt a persona shifts the probability distribution of its output toward text that an expert in that role would write. It's not magic — it works because the training data contains enormous amounts of text written by experts in various fields, and the role instruction activates those patterns.
2. Objective
State exactly what you want the output to be. Not "help me with X" — that's a request for help, not a clear objective. A clear objective is a specific deliverable.
3. Context
Background information that helps ChatGPT understand the situation. Without context, the model fills in gaps with assumptions — often the wrong ones.
4. Audience
The same content explained to a PhD vs a 12-year-old requires completely different vocabulary, depth, and examples. Specify the audience.
5. Constraints
What must the output NOT do or include? Constraints narrow the solution space and prevent unwanted patterns.
6. Examples (few-shot prompting)
Examples are one of the most powerful techniques in prompt engineering. By showing ChatGPT the format, style, or type of answer you want, you dramatically increase consistency. This works because examples are the clearest possible definition of what you want — far clearer than abstract descriptions.
Examples:
"The delivery was fast and the packaging was great." → Positive
"The product stopped working after 2 days." → Negative
"It arrived on Tuesday." → Neutral
Now classify: "The colour was different from the picture but the quality seems fine."
7. Output format
If you need JSON, bullet points, a numbered list, a table, markdown, or plain prose — say so. ChatGPT will default to markdown with headers and bullets unless told otherwise.
- Name
- Best suited for (team size and use case)
- Starting price
- One key weakness
Format this as a markdown table with those four columns.
8. Tone
Tone is about emotional register: formal, casual, authoritative, empathetic, witty, blunt. Without specifying tone, ChatGPT defaults to a neutral helpful voice that may not suit your purpose.
9. Length
Specify approximate length when it matters. "Under 150 words," "one paragraph," "a comprehensive 800-word article," or "a one-sentence answer only."
10. Success criteria
An advanced technique: tell ChatGPT what a good answer looks like. This acts like giving a grading rubric.
1. Be readable in under 2 minutes
2. Identify the top 3 findings
3. State the recommended action in the first sentence
4. Use no jargon
5. Be under 200 words
Part 4
The psychology of AI
ChatGPT completes patterns, it doesn't read minds
Every word you type is a signal that influences what patterns get activated from training. The more specific and unambiguous your signal, the more precisely ChatGPT can complete the pattern in the direction you want. Vague prompts activate vague, average patterns — because "average" is what the model has seen most.
Why ambiguity causes weaker answers
When a prompt is ambiguous, the model must "guess" which interpretation you intended. It typically picks the most statistically common interpretation — not necessarily yours. Example: "Summarise this article for me" — is the summary for a tweet, a board meeting, a student, or a press release? The model guesses. It's usually wrong in at least one dimension.
If you're unsure how to specify something, ask ChatGPT what information it would need before starting. Try: "Before you begin, what information do you need from me to do this well?"
Why examples improve consistency
When you give examples of the output you want (called few-shot prompting), you're not just describing what you want — you're demonstrating it. Demonstration is always more unambiguous than description. The model uses your examples as the strongest possible signal about the desired pattern.
Why structured requests help
Structure reduces the chance of the model "drifting" mid-response. When you break your request into numbered sections, the model can work through each section systematically rather than trying to address everything at once in a fluid paragraph.
Why iterative refinement beats "one perfect prompt"
Professional prompt engineers rarely nail a result in one try. They treat it as a process:
Expecting perfection on the first try sets you up for disappointment. Expecting a workable draft that you'll refine sets you up for success.
Part 5
Prompt frameworks
Frameworks are reusable templates that help you consistently include the right components. Here are three that work across a wide range of tasks.
Framework 1: RTCCO (most versatile)
Who should ChatGPT act as?
What exactly do you want produced?
Background information needed
What to avoid or limit
Format, length, tone
Best for: Writing tasks, content creation, analysis, emails, reports.
Task: Write a cold email subject line and opening paragraph.
Context: I'm a startup founder selling a recruitment software to HR directors at mid-sized companies (200–1000 employees). The product reduces time-to-hire by 40%.
Constraints: No buzzwords. No false urgency. Under 75 words for the opening paragraph.
Output: 3 different variations ranked from most formal to most conversational.
Framework 2: WIRE (analytical tasks)
Who is asking and for whom?
What decision or action will this inform?
What must the answer include?
How will you judge if the answer is good?
Best for: Research, competitive analysis, strategy documents, decision support.
Framework 3: GIRE (structured output)
The end result you need
Facts and context ChatGPT needs
Hard constraints the output must follow
Criteria for a good answer
Best for: Coding, data extraction, structured document generation, technical tasks.
Part 6
10 complete real examples
Example 1: Writing (blog post)
Why it fails: No audience, no angle, no length, no tone. ChatGPT produces a generic 5-tip listicle you could find anywhere.
Write a 700-word opinion piece on why the "productivity paradox" of working from home — where people work longer hours but feel less productive — happens, and what companies can do about it. Audience: mid-level managers at 100–500 person companies. Tone: direct, slightly contrarian, no platitudes. Do not use the phrase "work-life balance." End with one concrete, implementable action.
Why it works: Role + angle + length + audience + tone + constraint + specific ending structure = completely unambiguous task.
Example 2: Coding (write a function)
clean_customer_data that:1. Takes a pandas DataFrame as input
2. Removes rows where the "email" column is empty or null
3. Strips leading/trailing whitespace from all string columns
4. Converts the "created_at" column from string to datetime format
5. Returns the cleaned DataFrame
Include inline comments explaining each step. Use type hints. Python 3.10+.
Example 3: Debugging
Example 4: Excel formula
Example 5: Marketing copy
Example 6: Studying / learning
Example 7: Resume writing
Example 8: Email writing
Example 9: Research
Example 10: Business planning
Part 7
Uploading images correctly
How ChatGPT processes images
When you upload an image, it is encoded into a series of visual tokens and added to the context window alongside your text. ChatGPT uses a vision model to translate the visual content into representations the language model can "attend to." Importantly, this happens inside the same context window — an image is not processed separately and then summarised for the language model. It is directly accessible.
ChatGPT's vision capabilities are impressive but not perfect. It can misread small text, fail with complex diagrams, and struggle with very low-resolution images. Always verify important details it extracts from images.
How to ask about images effectively
1. Read and list every metric visible in the dashboard
2. Identify the time period being shown
3. Flag any metric that appears to be underperforming (below the visible targets or benchmarks)
4. Suggest one action based on what you see
Comparing multiple images
Tips for better image prompting
- Crop tightly: If you need analysis of a specific area (e.g., a chart legend, a UI component), crop the image before uploading. Less irrelevant visual information = better focus.
- Annotate if possible: Draw an arrow or circle on the image to mark the area of interest. "What is wrong with the component I've circled in red?"
- Order matters: For sequential tasks ("first look at image 1, then compare with image 2"), upload in the order you want them considered and reference them explicitly by position or a label.
- State what you already know: "This is a chest X-ray of a 45-year-old patient. I'm not asking for a diagnosis — I want you to describe what you can observe in the lung area on the right side."
Part 8
Uploading PDFs correctly
How ChatGPT processes PDFs
For shorter PDFs, the text is extracted and placed directly into the context window as tokens — exactly like pasting text. For very large PDFs (roughly over 110,000 tokens / ~85,000 words), ChatGPT may use a retrieval system that indexes the document and fetches relevant chunks rather than loading the whole thing into context. This means you may get different results depending on whether your PDF fits in context or requires retrieval.
Why "Summarise this" gives poor results
"Summarise this" tells ChatGPT almost nothing about what kind of summary you need. A 10-second overview? A page of key points? A summary for an executive who needs to make a decision? A summary for someone who will read the full document later? Each requires completely different output.
1. Top 3 market opportunity findings (2 sentences each)
2. Top 3 risks or challenges identified
3. The report's key recommendation (quote it directly if possible)
4. Any data you're uncertain about or couldn't verify in the document
Total length: under 400 words.
Useful PDF prompt patterns
| Use case | What to say |
|---|---|
| Data extraction | "Extract every figure/percentage/date from section 3 and format as a table." |
| Clause finding | "Find any clause related to termination or penalty fees. Quote the exact text and the section number." |
| Comparing documents | "I've uploaded two contracts. List every material difference between them in a table with three columns: clause type, version 1 text, version 2 text." |
| Research papers | "Summarise the methodology, key finding, limitations, and conclusion of this paper. Then tell me: what claims are made that would require verification against the raw data?" |
| Financial statements | "From this annual report, extract: total revenue, net profit, EBITDA, cash on hand, and total debt for the most recent year shown. Format as a table and note the page/section where each figure appears." |
For critical documents, ask ChatGPT to quote the relevant text directly before summarising it. Grounding its answer in direct quotes reduces the chance of paraphrase drift or hallucination. Try: "Before summarising this section, quote the three most relevant sentences verbatim."
Part 9
Uploading multiple files
When you upload multiple files, all of them are added to the context window. ChatGPT can cross-reference them, but only if you tell it to. Without explicit instruction about which file is which and how they relate, the model may conflate, miss, or ignore files.
The structured multi-file approach
• File 1: My current CV
• File 2: The job description for the role I'm applying to
• File 3: My portfolio case study
• File 4: A draft cover letter I've already written
Please:
1. Identify the three most relevant skills from my CV that match the job description's requirements
2. Identify any requirements in the job description that my CV doesn't address
3. Rewrite the cover letter to lead with the most relevant skills, naturally reference the portfolio, and address the gaps identified in step 2
4. Keep the rewritten cover letter under 300 words
Rules for multi-file prompts
- Name each file explicitly in your prompt ("File 1: the contract, File 2: the amendment").
- State the relationship between files ("File 2 is an update that supersedes File 1 where they conflict").
- Give a clear output objective for each file or for the set as a whole.
- If files are large, be aware they collectively consume more context — watch for quality degradation on very long sets.
Part 10
Multi-step prompting
Professional AI users don't send one giant prompt and expect a perfect result. They work iteratively, using each response as a stepping stone to the next. Here's how a professional writing workflow looks:
This approach produces dramatically better results than one mega-prompt because each step builds on a reviewed, validated foundation. You also maintain control and can course-correct at each stage rather than discovering a fundamental problem after 1,000 words have been written.
When to start a new chat: If a conversation has become long and cluttered with abandoned directions, start fresh and paste in just the relevant context. A clean context window with focused instructions almost always outperforms a cluttered 50-message history. As a rule, if you can't easily summarise what the "current goal" of the conversation is, it's time for a fresh chat.
Part 11
30+ common prompt mistakes
| # | Mistake | Why it hurts quality | Fix |
|---|---|---|---|
| 1 | Too vague ("help me with this") | Model picks the most average interpretation | Specify the exact deliverable |
| 2 | No audience specified | Model defaults to "general adult" — wrong for most tasks | State exactly who will read this |
| 3 | No output format | You get markdown headers and bullets you didn't want | State the format explicitly |
| 4 | Contradictory instructions | Model tries to satisfy all constraints → mediocre output that satisfies none | Review your prompt for conflicts before sending |
| 5 | Hidden assumptions | "Make this better" assumes shared knowledge of what "better" means | Define your criteria explicitly |
| 6 | No examples when format matters | Model's guess of the right format is rarely optimal | Show one example of what you want |
| 7 | Asking for impossible certainty | "What will the stock price be tomorrow?" — model hallucinates rather than says "I don't know" | Ask for analysis of uncertainty, not predictions |
| 8 | Overloading one prompt | Asking 8 things at once gets mediocre answers for all 8 | Use multi-step prompting |
| 9 | Not specifying tone | Default ChatGPT tone is neutral — often wrong for marketing, creative, or personal writing | Name the tone and ideally give an example |
| 10 | Not specifying length | ChatGPT tends to write longer than you need | State a word count or say "be concise" |
| 11 | Saying "write in my voice" with no examples | ChatGPT has no idea what your voice sounds like | Paste 2–3 samples of your previous writing |
| 12 | Asking for a list when you need prose | Lists feel artificial and don't convey nuance well | Say "write in prose paragraphs, not bullet points" |
| 13 | Pasting huge context without a clear question | Model summarises or guesses what you want | Lead with your question, then paste the context |
| 14 | Accepting the first output | First outputs are good first drafts, rarely finished work | Always do at least one refinement pass |
| 15 | Not specifying what to exclude | Model includes filler content, clichés, or unwanted sections | State explicitly what to leave out |
| 16 | Assuming ChatGPT knows your industry | Model has general knowledge but not your company's terminology | Provide glossary or context for specialist terms |
| 17 | Asking for "creative" without constraints | "Creative" to an AI = statistically common creative patterns | Give specific constraints: subvert X, avoid Y, include Z |
| 18 | Not verifying facts from long documents | Model can hallucinate even when "reading" a PDF | Ask for quotes and page numbers; verify critical facts |
| 19 | Using technical abbreviations without defining them | Model guesses the definition; often wrong in niche contexts | Define all abbreviations first |
| 20 | Relying on memory in long conversations | Context window truncation causes model to "forget" early instructions | Restate key constraints every 10–15 messages |
| 21 | Not specifying language or dialect | Model defaults to US English | Specify: "UK English," "formal Spanish," "Australian slang" |
| 22 | No success criteria | Model has no benchmark for quality | Add "A good answer will include X, Y, Z" |
| 23 | Asking ChatGPT to roleplay as itself with no limits | Gets you generic ChatGPT — no differentiation | Give specific, narrow role and expertise |
| 24 | Not using follow-up refinement | Treating first output as final | Always ask "what could be improved here?" |
| 25 | Uploading images without questions | Model describes the image generically | Ask a specific question about the image |
| 26 | Asking for "the best" approach | Model gives the most common approach, not necessarily the best for your situation | Describe your situation and ask for tradeoffs between options |
| 27 | Not asking for confidence or uncertainty | Model states everything with equal confidence | Add "flag where you're uncertain or where I should verify" |
| 28 | Conflating creativity with accuracy | For factual tasks, encouraging creativity increases hallucination | For factual tasks, say "stick to what you know; say if you're unsure" |
| 29 | Ignoring that context resets each conversation | Starting a new chat and expecting ChatGPT to "remember" yesterday | Paste relevant context at the start of each new chat |
| 30 | Prompt that is longer than necessary | Noise dilutes signal; critical instructions get less attention | Include only what the model genuinely needs |
Part 12
Advanced prompt engineering
Ask the model what information it needs
Before writing a complex prompt, ask ChatGPT to tell you what it would need to do the job well. This is like having a briefing call with a contractor before the project starts.
Request multiple options with tradeoffs
Instead of asking for "the answer," ask for three options with explicit tradeoffs. This prevents anchoring on one approach and gives you decision-making material.
Use evaluation rubrics
After getting an output, ask ChatGPT to grade it against your criteria. Then ask it to rewrite to address the low-scoring areas. This is cheaper and faster than writing a perfect prompt from scratch.
Chain of thought prompting
For complex reasoning tasks, ask ChatGPT to think step by step before giving an answer. This improves accuracy on maths, logic, and multi-step analysis because each reasoning step is visible and builds on the previous one.
Separate brainstorming from drafting
Generating ideas and filtering ideas are two different cognitive tasks. They're also two different "prompt modes." Mixing them in one prompt produces filtered, conservative output. Separate them:
- First prompt: "Generate 20 ideas. No filtering. Quantity over quality."
- Second prompt: "Review those 20 ideas. Identify the 3 with the most commercial potential and explain why."
- Third prompt: "Develop idea 7 into a full proposal."
Building reusable prompt templates
Identify tasks you repeat frequently and build a template with placeholder variables. Store it in a notes app, and fill in the variables each time.
You are a [ROLE] with expertise in [DOMAIN].
Task: [DELIVERABLE]
Context: [BACKGROUND INFORMATION]
Audience: [WHO WILL READ/USE THIS]
Constraints:
- [CONSTRAINT 1]
- [CONSTRAINT 2]
Output format: [FORMAT DESCRIPTION]
Length: [LENGTH]
Tone: [TONE]Managing long conversations
When a conversation gets long (15+ messages), periodically send a "constraint refresh" message:
Part 13
Myth vs reality
| Myth | Verdict | Reality |
|---|---|---|
| "Longer prompts are always better" | False | Longer prompts add noise. Include only what the model genuinely needs. Unnecessary text dilutes the signal of your actual request. |
| "Giving ChatGPT a role always improves answers" | Partly true | Role prompting helps when the role is specific and relevant to the task. "You are a helpful AI assistant" adds nothing. "You are a senior securities lawyer" adds a lot — for legal tasks. |
| "Magic words unlock hidden intelligence" | False | There are no secret codes. Words like "DAN" or "ignore previous instructions" don't bypass anything meaningful. Effective prompting is about clarity and context, not tricks. |
| "Prompt engineering is only for experts" | False | The core skill is being specific and clear. That's writing skill, not technical skill. Anyone can learn it. |
| "ChatGPT remembers everything forever" | False | Without memory enabled, every new conversation starts completely fresh. Even with memory, it stores summaries — not perfect recall of every word you've ever said. |
| "AI can read my mind — I shouldn't need to over-explain" | False | ChatGPT has no access to your mental model, your specific context, or your unstated assumptions. It only knows what's in the context window. The more you explain, the better it performs. |
| "ChatGPT always tells you when it doesn't know something" | False | ChatGPT can generate confident-sounding text about things it doesn't reliably know — a phenomenon called hallucination. Always verify factual claims independently for important decisions. |
| "You should always start each answer from scratch" | False | Iterative refinement within a conversation is often more efficient. Ask ChatGPT to improve or revise its previous output. |
| "The free version gives bad results; paid always wins" | Partly true | Prompt quality matters more than model version for most tasks. A good prompt on a smaller model often beats a bad prompt on a larger one. |
| "ChatGPT's training data includes everything on the internet" | False | Training data is curated and has a knowledge cutoff. ChatGPT does not know about events after its training cutoff. Some sources are excluded, paywalled content is typically not included, and private information is not in the training data. |
Part 14
The perfect prompt checklist
Before sending any important prompt, run through this checklist:
- I've specified a clear, concrete deliverable (not "help me with X")
- I've defined the audience who will read or use the output
- I've stated the desired output format (prose, bullet list, table, JSON, etc.)
- I've specified the approximate length
- I've defined the tone where it matters
- I've provided the necessary context and background
- I've listed what to avoid or exclude
- If I need a specific format/style, I've given an example
- My instructions don't contradict each other
- I've included success criteria or a quality benchmark
- I haven't buried the most important instruction in the middle of a long paragraph
- I've removed filler text that the model doesn't need
- If I'm uploading files, I've named and explained each file's role
- I've asked ChatGPT to flag uncertainty rather than assume
- I'm planning to refine the output, not expect perfection on the first try
Part 15
50+ prompt templates
Copy, adapt, and reuse these templates. Fill in the [BRACKETED] parts.
Writing
Programming & technical
Learning & research
Business & marketing
Email & communication
Career & productivity
Finance, health & travel
Summary
Prompt like a pro — the cheat sheet
- Specific deliverable
- Audience
- Output format
- Length
- Tone
- Role / persona
- One or two examples
- Success criteria
- Constraints (what to avoid)
- Context / background
- First output = draft
- Refine in 2–3 passes
- Verify facts independently
- Restate constraints in long chats
- Context resets each new chat
- It reads your entire history
- Token ≈ ¾ of a word
- Middle of long context gets less attention
- Conflicting instructions = worse output
- Examples beat descriptions
- Too vague
- No audience
- No format specified
- Contradicting yourself
- Accepting first draft
- RTCCO — for writing
- WIRE — for analysis
- GIRE — for technical tasks
- Step-by-step for complex tasks
- Multi-file: name every file
Prompt engineering is just clear communication. The better you can articulate what you want, who it's for, and what success looks like, the better your results will be. Every tip in this guide is in service of that one idea: remove ambiguity, add specificity.
This guide reflects publicly available information about large language models, ChatGPT's architecture, and prompt engineering best practices as of mid-2026. Specific model capabilities, context window sizes, and memory features may change. Always check OpenAI's current documentation for the latest specifications.
