Most "prompting guides" list 20 techniques. Most of them matter about 2% of the time. These four cover the situations where simple prompts consistently fall short.
1. XML Tags — When You Have Multiple Inputs
When you're feeding Claude a document plus instructions, or data plus a schema, XML tags make it unambiguous about what's what.
Without tags:
Summarize the following document and keep it under 100 words, use
bullet points. Here's the document: [500 lines of text]
With tags:
<instructions>
Summarize the document below. Bullet points. Under 100 words.
</instructions>
<document>
[500 lines of text]
</document>
The tagged version is structurally clear. Claude knows instructions from content. This matters a lot when the document itself contains instruction-like language — which it often does.
Use tags when you have:
- Long inputs where structure helps
- Multiple distinct inputs (document + schema, data + rules, etc.)
- Dynamic content being inserted programmatically into prompts
2. Chain of Thought — For Anything Involving Judgment
For tasks that involve real reasoning — tradeoffs, multi-step logic, anything where "it depends" is the honest answer — adding "think through this step by step before answering" reliably improves quality.
I'm deciding between these two database schemas for my app. Think
through the tradeoffs step by step before making a recommendation.
[Schema A]
[Schema B]
What this does: Claude externalizes its reasoning, which means you can see where it's going wrong if the conclusion doesn't make sense. You're not just getting an answer — you're getting an answer you can interrogate.
Skip it for simple tasks where you just need output fast. It adds tokens and time.
3. Few-Shot Examples — Show the Format You Want
Sometimes the fastest way to communicate what you want is to show an example of it.
Write a changelog entry in the same format as this example:
---
**Faster search** — Results now load 3x faster on large datasets.
Affects Pro and Enterprise plans. No action needed.
---
Feature: [your feature description]
Claude will match the length, structure, and tone more reliably than if you tried to describe those qualities in words. Two or three examples is usually enough for more complex formats.
This is particularly useful when you're generating many outputs that need to be consistent — API responses, email templates, product copy variants.
4. Ask Claude to Critique Its Own Output
Before you accept an output, ask Claude what's wrong with it.
Read what you just wrote. What are the three weakest parts?
How would you improve them?
Or more specifically:
Review your response above as a skeptical senior dev looking at a
junior's PR. What would you push back on?
Claude often knows where it cut corners or made assumptions — it just doesn't say so unless you ask. This catches issues before they become your problem.