Prompt Engineering 101: How to Write Prompts That Actually Work

Prompt engineering is the skill of writing instructions that help an AI model produce the result you want. A good prompt reduces ambiguity, adds the right context, and forces a useful output format.

Most "bad AI results" are not model failures. They're vague requests wearing a trench coat.

AI models are extremely sensitive to:

• missing context (audience, goal, constraints)

• unclear success criteria

• messy input (uncurated text, mixed instructions)

• output format (the model guesses what you want)

A good prompt is basically a product spec for one tiny task.

Use this structure:

• Goal: what you want to achieve

• Context: who it's for, what the situation is

• Constraints: limits, style, length, do's and don'ts

• Inputs: the material to use (paste it or describe it)

• Output format: bullets, table, JSON, steps, etc

• Quality bar: how to verify, what "good" looks like

Copy-paste template:

Goal: …

Context: …

Constraints: …

Input: …

Output format: …

Quality check: …

1) Ask for clarifying questions first

Template: "Before you answer, ask up to 5 clarifying questions. If you can proceed with assumptions, list them and continue."

2) Force structure

Template: "Output as: Summary (3 bullets), Main answer (sections with H2 headings), Checklist (10 items). Keep it concise."

3) Provide examples (few-shot)

Template: "Here is an example of the style I want: [Example]. Now produce the same style for: …"

4) Add constraints that prevent fluff

Template: "Avoid generic advice. Use concrete steps. No filler. If info is unknown, say 'Unknown' and suggest how to find it."

5) Make it compare options

Template: "Compare options A, B, C using criteria: cost, speed, quality, risk. Output as a table, then give a recommendation."

6) Work in passes

Template: "First produce a rough draft. Then do a second pass to improve clarity. Finally, output the final version only."

7) Add an evaluation rubric

Template: "After answering, score the output from 1–5 on: accuracy, completeness, clarity. Then revise to improve the weakest score."

8) Delimit your input clearly

Template: "Use only the text between and ."

9) Safety against hallucinations

Template: "If the answer is not supported by the provided input, say 'I don't have enough information' and ask for the missing data."

10) Tool-aware prompts

Template: "If you need external info, request a 'Search' step. If the task involves irreversible actions, ask for confirmation before proceeding."

A) Writing: blog outline

Goal: Create an SEO-friendly outline for an article.

Context: Audience is beginners, tone is practical.

Constraints: No fluff. Include unique angles, FAQs, and a checklist.

Output format: H1 + H2/H3 outline, then 8 FAQs, then a 10-point checklist.

B) Marketing: ad angles and variants

Goal: Generate ad angles and copy variants.

Context: Product is an AI transcription tool for podcasters.

Constraints: 5 angles, each with 3 headlines + 2 primary texts. No exaggerated claims.

Output format: Table with columns Angle, Headline, Primary text.

C) Research: summary with sources

Goal: Summarize the document for a decision maker.

Constraints: Only use the provided text. Quote short supporting snippets.

Output format: Executive summary (5 bullets), Key findings (with quotes), Open questions.

D) Coding: generate tests

Goal: Write unit tests for this function.

Constraints: Include edge cases and explain what each test covers.

Output format: Test file code first, then a short explanation.

E) Customer support: draft replies

Goal: Draft a customer support reply.

Context: Tone is calm, friendly, professional.

Constraints: Follow policy rules exactly. If user requests disallowed action, refuse and offer alternatives.

F) Presentations: create a slide plan

Goal: Create a 10-slide deck outline.

Context: Audience is executives, 5 minutes.

Constraints: One idea per slide. Include a clear narrative arc.

Output format: Table with Slide title, Key bullets, Speaker notes.

When output is wrong, usually one of these is missing:

• Goal is too broad

• Audience not specified

• Constraints absent (length, tone, forbidden claims)

• Input not delimited

• Format not forced

• Success criteria unclear

• You asked for "best" without criteria

• You mixed multiple tasks in one prompt

Fix by adding one missing piece at a time. Prompting is iterative engineering, not poetry.

If your site will publish lots of pages (use cases, alternatives, learn articles), prompts can help you keep quality consistent.

Good workflow:

1. Generate outline

2. Generate first draft

3. Add "unique insights" block manually or via a stricter prompt

4. Add FAQ

5. Add internal links (learn → use-case → alternatives)

And always do a human pass for claims, dates, and anything that sounds too confident.