Ask ChatGPT to “write an email to my landlord about the heating,” and you’ll get something polite, tidy, and useless. It won’t mention when the heat went out, what you’ve already tried, or what you want to happen next. It can’t. You never said.
That’s the whole problem with most AI prompts. The AI isn’t lazy, and it isn’t dumb. It’s guessing. Every detail you leave out is a blank it has to fill, and it fills each blank with the most average choice it can find. Average in, average out.
This post walks through a six-part checklist that removes the guessing. It’s called CROFTC: Context, Role, Objective, Form, Tone, and Conditions. We’ll look at what each part does, watch two vague prompts produce forgettable answers, then rewrite them with CROFTC and compare. No technical background needed. It works the same way in ChatGPT, Claude, or Gemini.
Why Vague Prompts Get Vague Answers
Think of the AI as a very fast, very well-read new hire who started this morning. It knows a lot about the world and nothing about you. Hand it a sticky note that says “fix the report,” and it will produce a fixed report, based on what most people usually mean by that. Whether that matches what you meant is luck.
A good manager wouldn’t hand over a sticky note. They’d explain the situation, say what “done” looks like, and mention the rules. That’s all CROFTC is: the questions a good manager answers before handing off a task. Other checklists exist (CO-STAR, CRAFT). They differ in the letters, not the idea.
Key insight: The quality of an answer is capped by how much guessing you leave to the AI. Each part of CROFTC removes one guess.
The Six Parts of CROFTC
| Part | The question it answers | What it sounds like |
|---|---|---|
| Context | What background does the AI need to understand the request? | “I rent an apartment in New York. The heat has been off for four days.” |
| Role | Who should the AI act as? | “You are an experienced property manager.” |
| Objective | What result do I want? | “Get a repair scheduled within 48 hours.” |
| Form | What shape should the answer take? | “A short email, under 150 words, with a subject line.” |
| Tone | What attitude should it use? | “Polite, calm, and firm.” |
| Conditions | What rules or limits must it follow? | “Include the date I reported it. No legal threats.” |
Context is the situation: who’s involved, what happened, what you’ve already tried. It’s the part people skip most and the part that matters most. Without it, the AI is writing for a stranger.
Role sets the vocabulary, priorities, and level of detail. “Explain this like a patient tutor” and “review this like a senior engineer” produce very different answers to the same question.
Objective is the outcome, not the task. “Write an email” is a task. “Get the heat fixed this week” is an outcome. Once the AI knows the outcome, it makes smarter choices about the task.
Form is the shape: a table, five bullet points, a 200-word summary, code with comments. Say it, or the AI picks “several paragraphs” by default.
Tone is the attitude: friendly, blunt, formal, playful, no fluff. Three words here often decide whether you can use the answer as-is.
Conditions are the rules: things it must include, things it must avoid, limits on length or scope. This is also where you can say “if anything is unclear, ask me before answering,” which stops a lot of wrong guesses before they happen.
Example 1: An Email to Your Landlord
The vague version
Write an email to my landlord about the heating.
Here’s what a typical answer looks like:
Subject: Heating Issue in My Apartment
Dear [Landlord’s Name],
I hope this email finds you well. I am writing to inform you that I have been experiencing issues with the heating in my apartment. It does not seem to be working properly, and I would appreciate it if you could look into this matter at your earliest convenience.
Thank you for your attention to this. Please let me know if you need any additional information.
Best regards,
[Your Name]
It’s polite, and it’s empty. No dates, no temperature, no deadline, no mention that you already reported it. The AI even had to invent “does not seem to be working properly,” because you never said what was wrong. Your landlord has no reason to act today.
The CROFTC version
Context: I rent an apartment in New York. The heat has been completely off since November 7 (four days ago). I texted my landlord on November 8 and got no reply. The indoor temperature is about 55°F.
Role: You are an experienced property manager who knows what makes landlords respond quickly.
Objective: Get a repair scheduled within 48 hours and create a written record of the problem.
Form: A short email, under 150 words, with a clear subject line.
Tone: Polite, calm, and firm. No threats and no apologies.
Conditions: Include the date I first reported the problem and the current temperature. Don't quote laws or mention lawyers. End by asking for a reply with a repair date.
The kind of answer this gets:
Subject: No heat since November 7 – repair needed by November 12
Hi Mr. Park,
The heat in apartment 4B has been completely off since November 7. I texted you about it on November 8 and haven’t heard back. The indoor temperature is now around 55°F.
Could you please have someone repair it by November 12? If a technician needs access, I’m home after 5 p.m. any day this week.
Please reply with the date and time you expect the repair to happen. Thank you.
Best,
Dana Lee, Apt 4B
Same AI, same request. The difference is that it stopped guessing. And look at what it added on its own: a deadline in the subject line and an offer of access times. Those came from the Objective. You asked to get the heat fixed, not just to send an email.
Example 2: Slow Python Code
The vague version
My Python script that compares two lists is really slow. How do I make it faster?
A typical answer is a list of general tips:
There are several ways to speed up Python code:
- Use built-in functions and libraries, which are implemented in C.
- Avoid unnecessary loops, especially nested loops.
- Consider using NumPy for numerical operations.
- Profile your code with cProfile to find bottlenecks.
If you share your code, I can give more specific advice!
None of that is wrong. It’s just a list you could find in any tutorial. And look at the last line: the AI is politely asking for the Context you skipped.
The CROFTC version
Context: I have a Python 3 script that finds the items that appear in both of two lists. Each list has about 200,000 email addresses. The function below takes about 10 minutes to run.
def find_common(a, b):
common = []
for x in a:
if x in b:
common.append(x)
return common
Role: You are a senior Python engineer reviewing a junior developer's code.
Objective: Make this run in seconds, and help me understand why the original was slow so I can spot the same mistake next time.
Form: First, a short paragraph explaining the cause. Then the fixed function with brief comments. Then one or two lines about any change in behavior.
Tone: Direct and practical, like a code review comment. Skip the compliments.
Conditions: Standard library only, no new packages. Keep the function name and signature the same. If the fix changes the output order or how duplicates are handled, say so clearly.
The kind of answer this gets:
The problem is
if x in b. Whenbis a list, Python checks it item by item, so every lookup scans up to 200,000 items. With 200,000 items ina, that’s up to 40 billion comparisons. A set is built for exactly this kind of lookup: checking whether something is in it takes about the same time no matter how big it gets.
def find_common(a, b):
b_set = set(b) # build a fast lookup table once
return [x for x in a if x in b_set] # each check is now near-instant
Behavior: the output is identical. It keeps the order of
aand keeps duplicates froma, just like the original. The only cost is the memory for the set, which is small for 200,000 strings.
The vague prompt got a tutorial. The CROFTC prompt got a fix, a reason, and a note about side effects. Because the Objective included “so I can spot the same mistake next time,” the AI explained the general idea (lists versus sets) instead of silently changing one line.
Key insight: The AI didn’t get smarter between the two prompts. You stopped making it guess. Whatever is in your head and not on the page, the AI has to invent.
Three Habits That Make Every Prompt Better
CROFTC gives you the structure. These three habits fill it in well.
1. Trade judgment words for numbers and names. “Make it shorter” becomes “cut it to under 100 words.” “Sound professional” becomes “no slang, no exclamation marks, no ‘I hope this finds you well.'” “Soon” becomes “by Friday.” Words like better, simple, and professional mean different things to different people, and the AI will pick one meaning. It may not be yours.
2. Break big tasks into steps. “Build me a budgeting app” is a wish, not a prompt. Split it up: “List the five features a minimal budgeting app needs.” Then: “Describe the data we need to store for feature 1.” Then: “Write the function that adds an expense.” Each step gets the AI’s full attention, and you can steer between steps instead of discovering at the end that step one went wrong. Outside code, the same trick is: outline first, then expand one section at a time.
3. Say what you mean, not what you assume. “Fix this” is perfectly clear to you, because you know what’s broken. To the AI it’s a mystery. “This function should return an empty list when the input is empty, but it crashes instead” is clear to both of you. A good test: could a smart stranger act on your sentence without asking a question? If not, the AI is guessing.
A Copy-and-Paste Template
Context: (What's the situation? Who's involved? What have you already tried?)
Role: (Who should the AI be? A tutor, a reviewer, an editor, a planner?)
Objective: (What does "done" look like?)
Form: (Table, bullet points, email, code with comments? How long?)
Tone: (Friendly, formal, blunt, encouraging?)
Conditions: (Must include… Must avoid… If anything is unclear, ask me before answering.)
You don’t need all six every time. A quick factual question needs an Objective and maybe a Form. The bigger the task, and the more the answer matters, the more parts you should fill in.
The Takeaway
Good prompting isn’t a trick or a secret phrase. It’s the habit of putting on the page what’s in your head: the situation, the person you need, the result you want, the shape it should take, the attitude, and the rules. Context, Role, Objective, Form, Tone, Conditions.
Try it now. Take the last prompt you sent to ChatGPT, find the two parts you skipped, and add them. Then compare the answers. If you get a before-and-after you like, share it in the comments. I’d love to see it.