Prompt guide

How to write better prompts for any AI

ChatGPT, Claude, and Gemini are only as good as the instructions you give them. This guide covers the five parts of a strong prompt, how each AI differs, and how to stop the model from inventing facts.

Updated for 2026 ModelsChatGPT · Claude · Gemini · Perplexity · v08-minute read

1. The five parts of a good prompt

Pushkin turns a rough note into these five parts. When all five are clear, the AI gives a specific answer instead of a generic one.

A

Role

Skip “Act as an expert.” Name the job, the field, and what they should judge the work by.

✓ Strong:

“Act as a Principal Database Reliability Engineer specializing in PostgreSQL high-availability failover architectures.”

B

The task

Say the one thing you want, with a direct verb such as design, audit, or rewrite. Do not leave the result vague.

✓ Strong:

“Conduct a static vulnerability audit on the following TypeScript handler, categorizing every risk by CVSS severity score.”

C

Context

Keep background, assumptions, and source material separate from the instructions. Headings help the AI tell your request apart from the material it should use.

D

Format

Say how the answer should look: JSON, a table, numbered steps, or a code diff. If you need only the data, say “No intro and no sign-off.”

E

What to avoid

Say what the AI must not do. Without that, it often fills the gap with generic filler.

- Do not introduce external npm dependencies beyond what is listed.
- Do not abbreviate code with “// ... rest of code” comments.
- Avoid marketing buzzwords (e.g. “seamless”, “unleash”, “game-changing”).

2. Keep the AI from inventing facts

A quick prompt often leaves out names, budgets, and other facts. Many rewrites invent those details. That makes the answer look finished and still be wrong.

How Pushkin leaves blanks

Pushkin does not invent facts. If you write “Write an investor update for our launch last month,” it leaves blanks you can fill in:

- Company Name: [Your Company / Product Name] - Reporting Period: [e.g. September 2026] - Key Metric Achieved: [e.g. $42k MRR, +34% MoM growth] - Top Customer Highlight: [Brief customer quote or logos closed] - Primary Blocker / Current Ask: [e.g. Hiring lead frontend engineer]

Fill in the blanks, then send. The AI will not have to guess the missing facts.

3. Each AI answers best to a different prompt

A prompt that works well in Claude can wander in ChatGPT, and the reverse is also true. Pushkin changes the structure to match the AI you are using.

ChatGPT & OpenAI o-Series

OpenAI

Works best with numbered steps and a fixed format. It may invent details if “do not” rules are buried in a long paragraph.

Write it this way:

• Use a numbered list for the steps.

• Say what a good answer looks like at the start.

• Name the format (for example, “Return only raw JSON matching this shape”).

Claude (Anthropic 3.5 / 3.7 Sonnet)

Anthropic

Follows tone, limits, and “do not” rules well when each part has a clear heading.

Write it this way:

• Split the prompt into labeled sections.

• Say how far the reasoning should go (“Before writing, compare X and Y”).

• Include a “What to do” list and a “What not to do” list.

Gemini (Google)

Google

Handles long context well when the task comes first, then short bullets.

Write it this way:

• Keep the instructions flat. Skip nested if-then chains.

• State the goal before the reference material.

Perplexity AI

Perplexity

Built for live search. Tell it how recent sources should be, what to cite, and what to do when sources disagree.

Write it this way:

• Name which sites to prefer or skip.

• Ask it to check dates on topics that change quickly.

v0 (Vercel) & Lovable (AI App Generators)

Code & UI

These tools stumble on abstract goals. Describe the screen: layout, parts, and what each state looks like.

Write it this way:

• Specify component layout (e.g. desktop 3-column grid, mobile drawer).

• Define interactive states (empty, loading, hover, error, success).

• List styling tokens: Tailwind CSS, Radix primitives, Lucide icons.

4. Example prompts

Two prompts Pushkin would write, one for code and one for a UI tool.

Code rewrite

Claude / GPT-4o
Role: Senior TypeScript & React Performance Specialist.
Objective: Refactor the provided component to eliminate hydration mismatches and minimize re-renders.

Source Code:
```tsx
[Paste your React component here]
```

Requirements:
1. Identify root causes of unnecessary re-renders (state scoping, memoization gaps).
2. Fix any client/server hydration disparities.
3. Keep external dependencies identical; do not add external state libraries.

Output Deliverable:
- Brief bulleted summary of root performance bottlenecks.
- Complete, copy-paste ready refactored component in TypeScript.
- Explanatory inline comments on critical fixes.

UI brief for v0 or Lovable

v0 / Lovable
Role: Modern Design Systems Engineer.
Objective: Build a responsive [Feature Name, e.g. Billing Settings Card] using Next.js, Tailwind CSS, and Lucide icons.

Visual & Layout Specs:
- Desktop: Two-column card layout with active subscription tier on left and usage metrics on right.
- Mobile: Single-column stacked layout with full-width primary CTA.
- Style: Monochromatic minimal palette, subtle border highlights, clean typography.

Interactive States to Include:
1. Active state: displays current plan, renewal date, and "Manage Billing" button.
2. Quota progress bar: displays percentage used with warning threshold at 80%.
3. Payment issue alert banner: shown conditionally when payment method has failed.

Constraints:
- Use Lucide icons.
- Ensure accessible ARIA labels for buttons and progress meters.

Apply these rules automatically with Pushkin.

You do not have to write this structure by hand. Use the Chrome extension or the web studio.