How to Prompt
Every conversation with AI starts with a prompt. Get the prompt right and everything after it gets easier.
Talk to it like a capable new coworker: naturally, concisely, conversationally. Think out loud, in plain words. You do not need special phrases.
Based on the prompting foundations in Anthropic’s Claude 101 and AI Fluency courses.
You get out what you put in
Ask a vague question, get a generic answer.
Give it who you are, what you want, and what good looks like, and the answer sharpens.
Nothing about the AI changes. What you feed it does.
Everything on this page, and honestly everything I teach, is one idea wearing different clothes: context. A prompt is context you type. A project is context you store. Your setup is context that’s always on. A second brain is context that compounds. SDS, below, is simply the recipe for arranging it.
Where are you working?
The Claude desktop app gives you more than one room to work in, and the most common problem I see in lessons is not a bad prompt. It is a good prompt asked in the room the task does not live in. Answer this question before you write anything.
Chat
A conversation. You steer, turn by turn.
- Quick questions, drafts, thinking out loud.
- Does not touch the files on your computer.
A Chat project
A workspace inside Chat that remembers you.
- Holds instructions and files you upload once, then keeps them for every conversation.
- Lives in your Claude account: it follows you to any device you sign in on.
A Cowork project
Points at a real folder on your computer.
- Claude reads, writes and organises the actual files in that folder.
- Lives on that computer, tied to that folder.
Two of those are called “projects”, and that is exactly where people get confused.
Here is the tell:
- A Cowork project shows its linked folder, and the app says the project stays on this computer.
- A Chat project has no folder. It only knows the files you upload to it.
If you made a “project” and Claude cannot see the files on your computer, you are in a Chat project. Nothing is broken. You are just in the other room.
Did a Cowork project disappear after a plan change or a new computer? Your work is not gone.
- A Cowork project is a link to a folder. The app lost the link, not the folder.
- Open Cowork, create a new project, choose Use a Folder, and point it at the same folder. Everything is back.
Keep your working folders in OneDrive or iCloud and the files themselves follow you between machines too.
- 1
Pick the room. Steering a quick question? Chat. A topic you return to with the same context? A Chat project. Real work on real files, handed over? A Cowork project.
- 2
Name the job. What exactly are you delegating, and what should come back? One job, one finished thing.
- 3
Then build the prompt. That is SDS, the rest of this page.
Room, job, prompt. That is the whole base case. Keep it by you: the one-page Base Case cheat sheet → Not sure whether a task is a Chat task or a Cowork task? The full lesson is Chat vs Cowork →
Watch context work
A real test, run live in lessons.
Recommend me a Greek mythology book.
The answer: the same five bestsellers everyone gets. Technically correct. Useless.
Recommend me a Greek mythology book. I loved Circe for the writing but found Mythos too jokey. I want something literary I can read in a week, and I already know the myths.
The answer: three titles you’ve never seen on a listicle, each with a reason tied to what you said.
Nothing about the AI changed. The context did. That second ask is just personalisation, and the three steps below are the recipe for writing it.
A good prompt, in three steps
Before your next conversation with Claude, consider: setting the stage (your role, objectives, and context), defining the task (what action you want Claude to take), and specifying rules (style, tone, and examples). Already met the 4D framework? SDS is its Description D made concrete: same move, sharper grip.
1. Set the stage
Give Claude the context only you have.
- Your role and objectives.
- What it should know.
2. Define the task
Name the action you want.
- Write, analyse, build, or something else.
- One job at a time.
3. Specify the rules
Say what good looks like.
- Style, tone, length.
- An example to copy, if you have one.
Set the stage
Give it the context only you have
AI cannot read your mind or your inbox.
A sentence of context is the cheapest win there is.
- What is your role, and what are your objectives?
- Is there context about your work that Claude should know?
Define the task
Name the one thing you want done
Vague in, vague out.
Be clear about the single action and the single output.
- What action do you want Claude to take?
- Write, analyse, build, or something else?
Specify the rules
Say what a good answer looks like
This is where most of the quality comes from.
Describe the shape of the answer you want.
- What style or tone do you want Claude to use?
- Are there examples you can attach to show what you are looking for?
An example prompt
All three steps in one real prompt. Each line is labelled with the step it does.
- 1Set the stage
I’m the marketing lead at an indie streaming startup, and we’re preparing an investor pitch deck for Series A investors.
- 2Define the task
Can you research the current state of the independent film streaming market and identify key trends, competitor positioning, and growth opportunities?
- 3Specify the rules
Use current web research with citations and structure it as a professional report of up to 5 pages, with an executive summary, market analysis, competitive landscape, and growth opportunities.
Let it interview you
Some ideas are too messy to type. Do not force the prompt: reverse the flow, and let Claude pull the context out of you.
Say this, word for word:
“I have an idea I cannot put into words yet. Interview me about it, one question at a time. When I say done, turn my answers into one clear prompt.”
- 1
It asks. One question at a time, and the questions surface context you would never have thought to volunteer.
- 2
You talk. Plain words, voice-to-text if you like. Messy is fine, messy is the point.
- 3
It writes the prompt. Your answers come back as one clean SDS prompt, ready to run.
Interview into SDS is how I plan my own builds: I speak, it asks, and my answers become the prompt (the sds-prompt skill does the same tidying on a voice note). The light version works in any prompt: add one line, “Before you answer, ask me two or three questions.”
When the first answer misses
Your first prompt is the start of a conversation, not a one-shot request. Treat the first answer as a draft, then steer. You have three moves.
Ask follow-ups
Build on what it gave you.
- “Expand on the second point.”
- “Make that more concise.”
Give feedback
Say what worked and what did not.
- “Good, but too formal.”
- “Cut the intro, keep the list.”
Redirect or restart
Steer it back, or start clean.
- “I meant X, not Y.”
- A fresh chat clears the context.
Common challenges, and how to fix them
When the answer is not quite right, it is usually one of these five. Each has a cause and a fix.
Too generic
Not enough context about your situation.
- Add your audience, role and constraints.
- Name specifics: “our enterprise client, second delay, keep it apologetic”.
Too long or too short
It is guessing at the right length.
- Give it a number: “two paragraphs”, “under 100 words”.
- Or lift the cap: “a full analysis, length no concern”.
The format is off
It got the what, but not the how.
- Paste an example of the format you want.
- Or spell out the structure: “a bold header, then bullets”.
Confident but false
It can state plausible but false facts, especially on niche detail.
- For anything that matters, verify the key facts.
- Ask for its sources, or turn on web search.
The tone is off
It defaults to helpful and professional.
- Name the tone: “more conversational”, “authoritative and formal”.
- Or give an example line in the voice you want.
The context ladder
Four steps. Each one means typing less and getting more.
- 1
Context you type. A good prompt: SDS, exactly as above. Too messy to type? Let it interview you first.
- 2
Context you store. A project that knows you: who you are, how you work, what good looks like. You stop retyping yourself.
- 3
Context that is always on. Your setup: the instructions and preferences Claude reads before every conversation, without being asked. The CLAUDE.md starter is the drop-in way to begin.
- 4
Context that compounds. A second brain your AI reads before every task. Context that gets richer every week. Ask me about this one in a lesson.
Most people live on step 1 forever. Every step up is the same lesson applied deeper. Step 2 starts at Claude 101 and Chat vs Cowork.
The three steps, remembered as SDS:
- Pick the room first: Chat to steer, a Chat project for a topic that keeps its context, a Cowork project for real files on your computer.
- Set the stage: give it the context only you have.
- Define the task: name the one thing you want done.
- Specify the rules: say what good looks like, and show an example.
- Let it interview you: when you cannot describe it, ask it to ask you.
- Keep going: follow up, give feedback, or restart clean.
- Troubleshoot: too generic, too long, format off, confident but false, or tone off, each has a fix.
- Everything is context: you get out what you put in.
Keep them by you: the one-page Base Case Cheat Sheet → (the four steps) and the SDS Cheat Sheet → (step three in detail)
You do not have to build the prompt by hand every time. The free sds-prompt skill → does the sorting for you: record a long voice note, or type a messy brain-dump, and it hands back one clean prompt built on these three steps, with anything it guessed clearly flagged. Your words, its structure. When you are ready for the bigger version of the same idea, meet the Translator →
New to the tools themselves? Meet them in Claude 101 → Point it at something real with Use Cases → Learn the framework behind good delegation in AI Fluency → And make every answer push back by default with the General Instructions template →