LLM best practices

1. Rule of Thumb

Don’t reinvent the wheel

If you asked LLM a question or to build something, assume a good human solution/template already exist (and fed into the LLM’s training data). You can follow up with a prompt like this:

Search in depth for existing open-source tools that perform the functions you described. 
Be very thorough with your search since I wish to give appropriate credit.

Note: that telling LLM to be thorough with the search (and why it should be thorough) usually makes it search more thorough.

Use XML Tags

Wrapping each content type in its own tag reduces misinterpretation and creates a clear hierarchy.

• Use consistent, descriptive tag names, for example: <context>...</context>, <task>...</task>, <instructions>...</instructions>.

• wrap links in <https://...>.

• Provide necessary resources directly in the prompt, for example:

Refer to the following Anthropic documentation on best practices of prompting: 

<https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices>

and ...

Formatting Output

Instead of saying, “Do not use markdown,” tell the model: “output the response in prose style.”

Formatting steering: If you want prose, write the prompt as prose.

• Using ALL CAPS sometimes help. For example: "OUTPUT ONLY IN ...!!!" often works to enforce formatting.
• Too much yelling can degrade performance on difficult task since it is distracting for model.

Please Ask

You can tell LLMs to ask you to make decision or supplement additional information (especially useful in Claude Code or other coding agent).

1. For example: Add “Ask me any clarifying questions before you begin” or “If you are uncertain, ask me.”
2. Meta-prompting: Paste the prompt and ask the model to critique and rewrite it. You can also use the prompt enhancer feature by many LLM providers.

---

2. Model Context

The “Lost in the Middle” Effect

In Long context, models attend best to the beginning and end of a prompt, but degrade in the middle.

• Long Text Placement: Put long documents and data near the top of the prompt, above your query and instructions.
• Quote-grounding: For needle-in-a-haystack tasks, tell the model to “quote the relevant parts of the documents first, then answer.”

Note: Both the user prompt + the model previous output count as context, the model sees the context in format generally something like: “<user_message>:…, <assistant_response>:…” etc.

Language Models are Few-Shot Learners

Providing examples is one of the single most reliable ways to steer format, tone, and structure.

• Show 3–5 examples of what “good” looks like.
• alternatively tell LLM to refer to a website, file, text etc.

---

3. Nuances

The Limits of Persona Prompting

While assigning a persona (e.g., “You are an expert coder”) helps structure tone and format, it does not unlock hidden capabilities.

• When “A Helpful Assistant” Is Not Really Helpful: shows personas help with alignment tasks (writing, role-play), but for pretraining-dependent tasks (math, coding), they can produce worse results.

Model Selection

• Use extended reasoning and State-of-the-Art (SOTA) models (e.g., Gemini 3.1 Pro Extended, Claude 4.8 Opus Max) for mission-critical tasks and searches. Use cheaper models for repeating similar, solved actions.

@misc{useful-llm-best-practices,
  title = {LLM best practices},
  author = {Yangyi Liu},
  year = {2026},
  url = {https://liuyangyi.com/posts/useful-llm-best-practices/}
}