This deck mirrors intro-gpt/prompts/. Pull it up alongside — every section in the slides has a deeper write-up there.
Five stages from your keystrokes to its first token. Skip this and you'll spend the rest of your career arguing with a black box.
Your prompt is broken into smaller units — tokens. Whitespace, punctuation, sub-words.
Uploaded documents and prior conversation are stitched in alongside your message.
The model weighs which parts of that context matter for the next token. This is where prompt phrasing pays off.
Response is produced token by token, conditioned on everything seen so far.
Output is structured according to your specifications — markdown, JSON, code blocks.
Most "the AI didn't listen" complaints are actually stage 03 problems: the relevant instruction was buried in a wall of context, and the model attended to something else.
| Feature | Claude | ChatGPT | Gemini | NotebookLM | Copilot |
|---|---|---|---|---|---|
| Document upload | PDFs, text, code | PDFs, images, data | PDFs, images, GDrive | PDFs, Google Docs | PDFs, OneDrive |
| Web search | via MCP | Yes | Yes | Yes | Yes |
| Context window | 200K | 128K | 2M | Document-based | 128K |
| File analysis | Yes | Yes | Yes | Deep analysis | Yes |
| Code execution | Yes (MCP) | Yes | Yes | No | Yes |
Context-window size is rarely the bottleneck — placement and phrasing inside it usually are. A 2M-token window does not exempt you from CRAFT.
# Basic "Summarize this research paper in 3 bullet points"
A request. No role. No structure. You can't predict what you'll get.
# Better "As a research scientist, summarize the key findings from this paper in 3 bullets, focusing on methodology and results"
Adds a role and a focus.
# Best "You are a research scientist reviewing for a journal. Summarize the PDF in 3 bullets: 1. Q & hypothesis 2. Methods & n 3. Findings & limits Format as a list."
Role · audience · scope · format. Reproducible.
Context · Role · Action · Format · Tone. Five fields. One mnemonic. The fastest way to make a prompt reproducible.
Provide background information. What project is this? Who is the audience? What's the upstream goal that this prompt serves?
"I'm preparing a grant proposal for NSF funding on AI in education…"
A role anchors the model to a register, a vocabulary, and a set of priors. "Grant writer" produces different prose than "tweet writer", even with the same task.
"Act as an experienced grant writer and education researcher…"
One unambiguous task. "Review" is not "rewrite". "Summarize" is not "expand". If you need three actions, number them.
"Review my draft introduction and suggest improvements…"
Markdown headers, bulleted lists, a comparison table, a JSON object, tracked-changes. The model will pick a default if you don't — and you usually won't like it.
"Provide feedback as tracked changes with explanations…"
Professional · constructive · skeptical · plain-English-for-a-policymaker. Tone is the one CRAFT field most people skip — and it's the difference between usable and rewriting from scratch.
"Professional, constructive, and encouraging."
Context: I'm preparing a grant proposal for NSF funding on AI in education. Role: Act as an experienced grant writer and education researcher. Action: Review my draft introduction and suggest improvements. Format: Provide feedback as tracked changes with explanations. Tone: Professional, constructive, and encouraging.
Every field constrains a different degree of freedom. The model has nowhere to drift.
A reusable CRAFT block becomes a custom instruction, a Project, or a prompt template.
System instructions · few-shot · chaining · multi-modal · web search. The vocabulary you need to recognize, even if you only use three of them this month.
# Project Context I'm a data scientist working on machine learning projects. Always provide Python code examples using scikit-learn and pandas. Include docstrings and type hints in all code. # Response Preferences - Be concise but thorough - Explain complex concepts with analogies - Always cite sources when making factual claims
It becomes a global rule applied to every prompt that follows — your defaults move up to the platform layer, not the chat.
I need to classify customer feedback. Here are examples: "The product arrived damaged" → Category: Shipping Issue "Can't log into my account" → Category: Technical Support "Love the new features!" → Category: Positive Feedback Now classify these: 1. "The app keeps crashing on startup" 2. "Best purchase I've made this year" 3. "Package was left in the rain"
3–5 examples is the sweet spot. Cover the corners of the label space, not just the easy cases.
Few-shot is how you teach a model your local ICD coding shorthand, your survey response taxonomy, or your team's review rubric — without fine-tuning.
Each turn corrects course. You're trading one undirected lottery ticket for four checkpoints where you can intervene.
I've uploaded: 1. A screenshot of my dashboard 2. The underlying data in CSV 3. Our brand guidelines PDF Create a redesigned dashboard: - improve data viz - adhere to our brand colors - highlight the data-dict KPIs
A single prompt that sees pixels, rows, and rules at once.
Search for the latest research on the public health benefits of vaccination published in 2024. Focus on: - Top conferences (AHA, ASPPH) - mRNA, Bird Flu, COVID Summarize the top 5 with links.
Pair retrieval with a citation requirement — or you get confident summaries with no source trail.
The mistakes that show up in every workshop, every cohort, every project — and the small rewrites that fix them.
"Make this better."
"Improve this abstract by making it more concise (under 250 words), adding keywords, and ensuring it follows the journal's structure: background, methods, results, conclusions."
Uploading 50 documents without guidance.
"Focus on documents 1–3 which contain the methodology. Ignore the appendices."
"Fix the usual issues."
"Check for: passive voice, sentences over 25 words, undefined acronyms, and missing Oxford commas."
"Summarize this."
"Create an executive summary with: a 3-sentence overview, 5 key points as bullets, 1 paragraph on implications, formatted with markdown headers."
Pull up a chat window. Pick a task from your week. Write it CRAFT-style. Compare it to the version you'd have typed an hour ago.