Writing a Great Prompt

Most “bad prompts” aren’t wrong—they’re just missing parts.

When someone says, “ChatGPT is inconsistent,” what they often mean is: I gave it a one-liner and hoped it would read my mind. (Relatable.)

A great prompt doesn’t rely on mind-reading. It gives the model four ingredients:

1. Context — who/what/why this is for
2. Instruction — the job to do (and rules)
3. Input Data — the raw material to work with
4. Output Indicator — what the final answer should look like

If you master this anatomy, you can prompt like a pro whether you’re writing Python, drafting legal language, or sending a sales follow-up.

1) Context: Put the Model in the Right “Room"

Context is the background the model needs to behave appropriately.

Examples of context:

• Who the audience is (VP of Sales vs new intern)
• Your goal (inform, persuade, de-escalate, debug)
• Constraints (brand voice, legal risk, company policy)

For AI engineers: context can include the stack, runtime, libraries, and failure mode.

2) Instruction: Give a Clear Job (Not a Vibe)

Instruction is what you actually want done.

Strong instructions:

• Start with an action verb (“summarize,” “draft,” “refactor,” “compare”)
• Add requirements (“keep it under 120 words,” “include 3 options,” “cite sources”)
• Include “don’ts” sparingly (“don’t invent metrics,” “don’t assume access to databases”)

3) Input Data: Provide the Ingredients

Input data is the text, bullet points, code, table, or transcript you want the model to transform.

This is where many prompts fail. People ask:

“Write my Q4 update”

…but don’t provide:

• accomplishments
• metrics
• blockers
• next steps

No inputs = the model improvises.

4) Output Indicator: Specify the Shape of the Answer

Output indicators are format instructions that force clarity.

Examples:

• “Return a table with columns: Risk, Impact, Mitigation”
• “Output valid JSON with fields: title, summary, action_items”
• “Use exactly 5 bullets, each under 12 words”

This is how you get consistency across teams.

Example 1: Marketing + Sales (Reliable Copy, Not Random Copy)

Here’s a prompt that uses all four parts.

CONTEXT:
You are a B2B copywriter for JoinAISchool. Audience is a VP of Sales who wants their team to use AI responsibly.

INSTRUCTION:
Write a short outbound email (90–110 words) inviting them to a 4-week prompt engineering program.
Tone: confident, helpful, not hype.
Include: subject line + one clear CTA.

INPUT DATA:

* Course: Prompt Engineering for teams (hands-on)
* Outcome: better emails, faster research, safer workflows
* Proof point: “Used by cross-functional teams (engineering + revenue)”

OUTPUT INDICATOR:
Return exactly:

1. Subject line
2. Email body

Why it works: the model knows the audience, the task, the ingredients, and the exact format.

Example 2: AI Engineer (Debugging With Boundaries)

Now let’s do a technical prompt that’s structured and safe.

CONTEXT:
You are a senior Python engineer helping debug a FastAPI endpoint.
Environment: Python 3.11, pytest, pydantic v2.

INSTRUCTION:
Identify the bug and propose the minimal fix. Then add one regression test.
Do not change function signatures unless necessary.

INPUT DATA:
<stack trace here>
<relevant code here>

OUTPUT INDICATOR:
Return:
- Root cause (3 bullets)
- Patch (code block)
- New test (code block)

Again: context + instruction + input + format = fewer surprises.

Takeaway

A great prompt isn’t long—it’s complete.

When your AI output misses the mark, don’t panic and start over. Use the anatomy checklist:

• Context: Who is this for and what’s the situation?
• Instruction: What should it do (and what should it avoid)?
• Input Data: What info should it use to stay grounded?
• Output Indicator: What format makes “good” undeniable?

Get these four right, and your prompts will feel less like gambling—and more like engineering.