DEV Community

Nova
Nova

Posted on

5 Prompt Engineering Mistakes That Are Costing You Time (And How to Fix Them)

#ai #promptengineering #productivity #beginners

If you feel like you’re “good at prompting” but still losing time wrestling with ChatGPT, it’s usually not because you need more fancy frameworks.

It’s because of a few boring, repeatable prompt engineering mistakes that quietly create rework: unclear inputs, unclear outputs, and no feedback loop.

Below are 5 prompt engineering mistakes I see constantly (including in my own early prompts), plus practical fixes you can copy/paste today. Think of these as ChatGPT tips for writing better prompts that get you to a usable answer faster.

1) Mistake: Vague instructions ("do the thing")

Vagueness forces the model to guess your goal, audience, constraints, and success criteria. The output can be “technically correct” and still useless.

Fix: Add context + goal + constraints + definition of done.

Before (vague):

Write documentation for my API.
Enter fullscreen mode Exit fullscreen mode

After (better prompt):

You are a senior developer writing docs for external users.

Context: I have a REST API for creating invoices.
Goal: Produce a single-page "Quickstart" doc for first-time users.
Audience: JavaScript devs.
Constraints:
- Use concise language.
- Include curl examples and a Node.js fetch example.
- Assume base URL is https://api.example.com.
Definition of done:
- Sections: Auth, Create invoice, Common errors, Next steps.

Here is the endpoint spec:
POST /v1/invoices
Body: { customerId: string, items: [{ sku: string, qty: number }] }
Returns: { id: string, status: "draft" | "sent" }
Enter fullscreen mode Exit fullscreen mode

Why this saves time: you’re buying clarity up front instead of paying for 3–5 rounds of “not what I meant.”

2) Mistake: Asking for an answer without giving examples

If you want a specific style, structure, or level of depth, show it. Otherwise the model will pick a “default” voice that might not match your needs.

Fix: Provide one tiny example (even a partial one) and ask it to match.

Before (no examples):

Rewrite this email to be more professional.

[PASTE EMAIL]
Enter fullscreen mode Exit fullscreen mode

After (better prompt):

Rewrite the email to sound calm, direct, and professional.

Match this style:
- Short paragraphs
- One clear ask
- No exclamation marks

Example tone:
"Hi Sam — quick update: we can ship on Friday if we get approval by Wednesday. Can you confirm by EOD?"

Now rewrite this email:
[PASTE EMAIL]
Enter fullscreen mode Exit fullscreen mode

Pro tip: When you’re iterating on a pattern, keep a small “prompt snippet library” of examples you like. Reuse them.

3) Mistake: Not specifying the output format

This is the fastest way to end up with a wall of text when you needed a checklist, JSON, or a table you can paste into a ticket.

Fix: Tell the model exactly what to return (and what not to return).

Before (no format spec):

Give me a plan to improve our onboarding.
Enter fullscreen mode Exit fullscreen mode

After (better prompt):

Create an onboarding improvement plan.

Output format:
1) A table with columns: "Issue", "Impact", "Fix", "Owner", "Effort (S/M/L)", "Success metric".
2) Then a prioritized checklist (max 10 items).

Constraints:
- Keep it under 250 words.
- Assume a 4-person startup team.
- Focus on the first 14 days of a new user.

Do not include generic advice like "listen to users" unless it’s tied to a specific action.
Enter fullscreen mode Exit fullscreen mode

Why this matters: models are great at generating text, but you’re usually trying to generate artifacts (tickets, specs, schemas, steps).

4) Mistake: Ignoring temperature (and then blaming the model)

“Temperature” is basically the creativity knob. Higher temperature = more variety and risk. Lower temperature = more deterministic and literal.

If you’re using an API, you can set it explicitly. If you’re using ChatGPT’s UI, you can still simulate it by telling the model what mode you want.

Fix: Decide what you’re doing, then set the mode:

  • For factual / structured work: low temperature (or “be strict, don’t invent”).
  • For brainstorming / naming / ideation: higher temperature (or “give 20 diverse options”).

Before (mode mismatch):

List potential root causes for this production bug and propose fixes.

Logs:
[PASTE LOGS]
Enter fullscreen mode Exit fullscreen mode

After (better prompt, low-temp behavior):

Act as a production incident responder.

Mode: Be conservative and evidence-based. If you’re unsure, say "unknown" and list what data you’d need.

Task:
1) Extract 5 concrete signals from the logs.
2) List up to 3 plausible root causes (ranked) and why.
3) For each root cause, give one verification step and one safe mitigation.

Logs:
[PASTE LOGS]
Enter fullscreen mode Exit fullscreen mode

And for brainstorming (high-temp behavior):

Brainstorm 25 punchy names for a developer tool that checks pull requests for security issues.

Mode: Be creative, avoid generic names, include a mix of one-word and two-word options.
Output: bullet list only.
Enter fullscreen mode Exit fullscreen mode

ChatGPT tip: If you want consistent results, ask it to “use a strict, repeatable process” and to avoid adding extra assumptions.

5) Mistake: Treating prompts as one-shot instead of iterative

Most real tasks aren’t “one prompt, one answer.” They’re specifying, checking, and tightening. If you skip the loop, you’ll waste time editing the output manually.

Fix: Build iteration into the prompt: first clarify, then draft, then self-check.

Before (one-shot):

Write a PRD for a new feature: "Saved searches".
Enter fullscreen mode Exit fullscreen mode

After (better prompt with iteration):

You are a product-minded engineer.

We’re writing a PRD for: "Saved searches" in a B2B dashboard.

Iteration loop:
Step 1: Ask me up to 7 clarifying questions (only questions).
Step 2: After I answer, write the PRD with sections: Problem, Goals/Non-goals, User stories, UX notes, Edge cases, Analytics, Rollout.
Step 3: Run a self-review checklist:
- Are goals measurable?
- Any missing edge cases?
- Any ambiguous terms?
Output the checklist results as bullets.

Start with Step 1.
Enter fullscreen mode Exit fullscreen mode

Why this is faster: you front-load alignment and force a quality pass, instead of doing “prompt ping-pong” for half an hour.

A quick copy/paste template for better prompts

When in doubt, use this skeleton:

Role: [who the model should be]
Context: [what’s going on]
Goal: [what you want]
Constraints: [time, length, tools, assumptions]
Output format: [exact structure]
Examples: [1 small example]
Quality bar: [how to judge success]

Task: [the actual request]
Enter fullscreen mode Exit fullscreen mode

That template alone eliminates most prompt engineering mistakes.

Wrap-up

Better prompting isn’t about magic words. It’s about reducing ambiguity: clear intent, clear format, and a built-in feedback loop. Do that and you’ll spend less time fighting the model and more time shipping.

Want more patterns like this? I put together a free Prompt Engineering Cheatsheet sample — grab it here: https://getnovapress.gumroad.com/l/prompt-cheatsheet

Top comments (0)