Prompt engineering cheatsheet.

Anatomy of a good prompt

  1. Role / context: who/what the LLM is.
  2. Task: what to do.
  3. Input: data to operate on.
  4. Format: structure of output.
  5. Constraints: tone, length, things to avoid.
  6. Examples: optional but powerful.
You are an expert technical writer.

Summarize the following article in 3 bullet points.
Each bullet must be one sentence and under 20 words.
Use neutral tone. Don't include opinions.

Article:
{text}

Output as JSON: {"summary": ["...", "...", "..."]}

System message

Used for stable behavior across calls. Keep concise:

You are a helpful coding assistant. Answer in concise Python.
Always include type hints.
If unsure, say so explicitly.

Zero-shot

Translate "Hello" to Spanish.

Few-shot

Translate English to Spanish.

en: Hello
es: Hola

en: Goodbye
es: Adiós

en: Thank you
es:

3-5 examples typically enough. Match desired output exactly.

Chain-of-thought (CoT)

Q: There are 3 apples. I eat 1. How many remain?
A: Let's think step by step. Started with 3. Ate 1. So 3-1=2.
The answer is 2.

Often unnecessary with reasoning models (o-series, Claude with extended thinking). Use for non-reasoning models.

Self-consistency

Ask multiple times, vote. Costs more, more accurate for hard tasks.

Tree of thoughts

Branch and prune reasoning paths. Complex; consider just using a reasoning model.

Structured input/output

Input as JSON:
{"name": "Alice", "age": 30}

Output as JSON only:
{"summary": "<summary>"}

Do not include explanation. Do not wrap in markdown.

For reliable JSON: use the API’s JSON mode / tool calling.

Constraints

- Do not invent facts.
- If you don't know, say "I don't know."
- Cite sources from <context>.
- Keep response under 200 words.
- Use Markdown headings.

Persona

You are a senior software engineer with 20 years of Linux experience.
You are blunt and concise.

Subtle effect. Don’t rely on persona alone for correctness.

XML tags (Anthropic prefers)

<task>Summarize the article</task>
<article>
...
</article>
<format>3 bullet points</format>

Clarifies structure for Claude especially.

Markdown headings

# Article

# Task

Summarize ...

Most LLMs respect markdown structure.

Avoiding hallucinations

  • Provide source material; ask LLM to only use it.
  • “If unsure, say I don’t know.”
  • Validate output against schema / facts.
Use ONLY the information in <context> below.
If the answer isn't there, respond with "NOT_FOUND".

Long context

  • Put critical instructions at start AND end.
  • Use clear sectioning.
  • LLMs sometimes ignore middle. Test.

Refusal mitigation

Frame requests legitimately. “Educational” / “for security research” can help if genuine.

Decoding from output

For when LLM is verbose:

Respond with ONLY the JSON object.
Do not include "```json" or any commentary.

Few-shot for format

Even ZERO content examples can show desired format:

Examples of output format:
{"name": "...", "score": 95}
{"name": "...", "score": 80}

Now process: <data>

Eval prompts

For evaluating other LLM outputs:

Rate the following response 1-5 on helpfulness, accuracy, and tone.
Return JSON: {"helpful": N, "accurate": N, "tone": N, "reason": "..."}

Temperature for prompt design

  • Test prompts at low temp (0-0.3) to see “default” behavior.
  • Use higher temp for variety / brainstorming.

Iterative refinement

Treat prompts as software:

  1. Define test cases.
  2. Try prompt.
  3. Note failures.
  4. Adjust.
  5. Re-test.
  6. Version-control.

Anti-patterns

  • “Be creative” without specifics → drift.
  • Negative-only constraints (“don’t do X”) → unclear what TO do.
  • 50 instructions in one prompt → some ignored.
  • Single long prompt vs structured conversation → conversation often clearer.
  • Trust LLM math without verification.

Reasoning models (o1, o3, Claude with extended thinking)

  • Less prompt engineering needed.
  • Don’t use CoT — model does it internally.
  • Give task + constraints; let it reason.
Solve this problem step by step. Output only the final answer in <answer> tags.

Problem: ...

Reusable prompt templates

PROMPT = """
You are {role}.

Task: {task}

Input:
{input}

Output format:
{format}

Constraints:
- {c1}
- {c2}
"""

Use Jinja2 / f-strings.

Common mistakes

  • Vague prompts → vague outputs.
  • Mixing examples with current task without separators.
  • Asking LLM to “be concise” without specifying length.
  • Forgetting that LLM continues your text.
  • Not testing on edge cases.

Read this next

If you want my prompt library, it’s at rajpoot.dev .

Building something AI-, backend-, or data-heavy and want a second pair of eyes? I do consulting and freelance work — see my projects and ways to reach me at rajpoot.dev .