Before you start prompting, you need an honest mental model of what you're working with. Claude is powerful, but treating it like magic will get you into trouble fast.
What Claude does well
Writing and rewriting: emails, summaries, documentation
Explaining complex things in plain language, at your level
Generating code: scripts, formulas, automations
Breaking down problems by turning vague goals into steps
Reviewing and improving drafts, scripts, and plans
What Claude doesn't do well
Real-time information: it has a knowledge cutoff date
Precise math: always verify numbers it produces
Knowing your context: it knows nothing about SRI unless you tell it
Being right every time: it can be confidently wrong
Memory between sessions: each conversation starts fresh
The most important concept: hallucinations
Claude sometimes generates information that sounds completely correct but is wrong. It's not lying. It's predicting what a plausible answer looks like based on patterns in its training. This is called a hallucination.
Where it happens most
Specific names, dates, URLs, API documentation, exact numbers, legal or compliance details
How to catch it
Ask Claude to explain its reasoning. Check any fact that matters. Never paste a URL it generates without verifying it exists.
Your job vs. Claude's job
You, The Decision Maker
Define what to build, why it matters, and what the constraints are. Know your systems, your goal, and what success looks like. These decisions belong to you. Claude can't make them for you.
Claude, The Executor
Takes your direction and figures out how to carry it out. Writes the code, structures the logic, formats the output. Highly capable, but only as good as the direction you give it.
Key insight: Claude doesn't know your company, your systems, or your role unless you tell it. Every session starts blank. The quality of what you get out depends entirely on what you put in. Think before you type. The more clearly you've thought through what you want, the better the result.
Most weak outputs come from weak prompts, not a weak AI. There's a simple formula that makes a dramatic difference.
The Prompt Formula
Context+Action+Format
Context: Who you are, what you're working on, and what the goal is. Action: Exactly what you want Claude to do. Be specific. Format: How you want the output: a list, a script, step-by-step, a table, or a short paragraph.
Bad prompt vs. good prompt, same task
Write me a script to export from AirtableWeak
No context, no output format specified, and ambiguous. Claude has to guess everything.
I'm an admin at Sorting Robotics. I manage our inventory in Airtable and need to export a list of all parts with a stock count below 10 into a CSV file so I can share it with our purchasing team. Write me a Python script that connects to our Airtable base, filters for rows where "Stock Count" is less than 10, and saves them to a file called low-stock.csv on my desktop. Use comments to explain each step in plain English.Strong
Role + system + goal + exact output + file name + plain-English comments requested. Claude has everything it needs.
More examples by role
Summarize this email threadWeak
I'm in sales at Sorting Robotics. Summarize this email thread in 3 bullet points covering what the customer asked for, what we promised, and what's still open. [paste thread]Strong
Help me track machine downtimeWeak
I'm a technician tracking machine downtime on the Stardust line. I have a spreadsheet with columns: Date, Machine ID, Downtime Minutes, Cause. Write me a formula for Google Sheets that totals downtime per machine for the current month.Strong
What if you don't know the implementation yet? The strong Airtable example above works because the writer already knew the details: CSV format, filtered below 10, saved to the desktop. But if all you know is "write me a script to export from Airtable" and you're not sure of the specifics, don't force it. Use /grill-my-plan first. It asks you questions until the picture is clear. Then you'll have everything you need to write a strong prompt.
Rule of thumb: If you could hand your prompt to a new employee on their first day and they'd know exactly what to do, it's a good prompt. If they'd have to ask clarifying questions, add those answers to your prompt first.
Step 2 of 5
Step 3Refining your results
Your first answer is never your last
Claude's first response is a starting point, not a final product. The real power comes from the back-and-forth: redirecting, refining, and building on what it gives you.
Four ways to improve an output
Correct directly. Tell Claude what's wrong and what you want instead. Don't just say "that's wrong." Say: "That's not right. The stock count column is called 'Units on Hand', not 'Stock Count'. Fix the script."
Ask for refinement. Change the scope, tone, or level of detail. "Make this shorter." "Use simpler language. I'm not a developer." "Add more detail to the error handling section."
Build on it. Extend what Claude gave you rather than starting over. "Good. Now add a step that sends me an email when the script finishes." "Now do the same thing but for the second table."
Ask it to explain. Before running any code, ask Claude to walk you through it. "Explain what this script does, line by line, in plain English. No technical terms."
Keep the conversation going. Don't start over. Each new conversation loses all context. When you're iterating on a script or document, stay in the same thread. Claude remembers everything you've discussed in that session.
What to do when Claude is stuck or going in circles
Restate your goal from scratch. Sometimes a long thread confuses the context. Say: "Let me restart. My goal is [X]. Ignore everything above and start fresh."
Break the problem smaller. If Claude keeps getting the same part wrong, isolate that part. "Forget the full script for now. Just show me how to read a single row from Airtable."
Ask it to verify its own work."Before I run this, review the script yourself and tell me if you see any issues or things that might not work."
Important: Claude will often say "You're right, I apologize" and fix something, then introduce a new problem in the same response. Always re-read the full output after a correction, not just the part that changed.
Step 3 of 5
Step 4Scoping your work
The most important skill: starting small
The most common mistake beginners make is asking Claude to build everything at once. Big requests produce fragile, hard-to-debug results. Small requests build toward something that actually works.
🚫
The mistake
"Build me an automation that pulls our Airtable inventory, checks HubSpot for pending orders, flags anything low-stock, and sends a Slack message to the purchasing channel every Monday morning."
✓
The right approach
Start with step 1 only: "Read my Airtable inventory and print all rows where stock is below 10." Get that working. Then add the next piece.
How to break any project into steps
Name the end goal in one sentence. If you can't, it's too vague. "I want to automatically notify purchasing when stock drops below 10 units."
List what data is involved: where does it come from, and where does it go? "Data comes from Airtable. Notification goes to Slack."
Build the first piece only: reading the data, before any writing or sending.
Verify it works before adding anything else. A script that reads correctly is a foundation. A script that reads, writes, and sends all at once is just a guess.
Add one piece at a time and test after each addition.
Use your /grill-my-plan skill
SRI Skill
You have a custom Claude skill installed called /grill-my-plan. It's designed specifically to help you scope a project before building it. It asks you questions one at a time about your goal, the data involved, the risks, and what success looks like, then produces a clear plan you can hand back to Claude to build from.
Use it any time you're starting something new and aren't sure how to break it down, or when you only know the end goal but not the implementation details. If all you can say is "I want to export from Airtable" or "I want to automate my follow-ups", that's exactly when to run this first. It works especially well for non-technical users because it never asks you to make technical decisions. It handles those for you.
Type: /grill-my-plan and then describe your project idea in plain language
A project that takes 5 focused prompts is better than one that takes 30 corrections. Smaller scope = faster wins = more confidence to keep building.
Step 4 of 5
Step 5Assignment
Put it into practice
Three tasks. Each one builds on the last. By the end you'll have a scoped plan for your first real project and a clear sense of what makes a prompt actually work.
Your goal this session
Go from vague to specific, then scope a real project
You're going to experience firsthand what a weak prompt produces versus a strong one. Then you'll use the /grill-my-plan skill to start turning your own project idea into something buildable.
Complete all three tasks
Write a weak prompt and notice what you get
Think of a repetitive task you do at work. Write a one-sentence prompt asking Claude to help with it: vague, no context, no format. Submit it and read the response. Notice how generic or unusable it is.
Try It
Example weak prompts to model yours on:
"Help me organize my files"
"Write a report about inventory"
"Automate my email follow-ups"
Rewrite it using Context + Action + Format
Take the same task and rewrite the prompt properly. Add your role, what system you're working with, exactly what you want done, and how you want the output formatted. Compare the two responses.
Compare
Use this structure:
"I'm a [role] at Sorting Robotics. I work with [system]. I need to [specific action]. Give me [specific output format] so I can [end goal]."
Use /grill-my-plan to scope your first project
Think of a real task at work you'd like to automate or simplify. Open Claude Code, type /grill-my-plan, and describe your idea in plain language. Let it guide you through the questions. Answer honestly, even if you're unsure. By the end you'll have a scoped plan.
Build
How to start:
/grill-my-plan
Then describe your idea: "I want to automate [task]. I work in [department] and currently do this manually by [how you do it now]."
0 of 3 tasks completed
✓ Module 2 complete
You know what makes a good prompt, how to refine Claude's output, and how to scope a project before building it. Next up is Security & AI Risks, the most important module before you touch any real system.
Save your /grill-my-plan output. You'll need it in Week 2 when you start building. Copy it somewhere you can find it: a notes app, a Google Doc, anywhere.
