Prompt Engineering

One of the most practical prompt engineering skills is getting Claude to produce output in an exact format — JSON you can parse, a markdown table you can paste into a doc, a CSV ready to import, HTML you can drop into a page. Claude can do all of these reliably, but the prompting matters. This chapter covers the patterns that work and the failure modes to watch for.

The Golden Rule: Specify Format Last

As established in Chapter 1, the end of your message carries the most weight. Format instructions belong at the end of your prompt — after the task, after the context, after the constraints. If you put them at the beginning, Claude may drift by the time it gets to generating the output.

Getting Clean JSON

JSON is the most commonly requested structured format — it's the lingua franca of pipelines, scripts, and APIs. The challenge is that Claude defaults to wrapping code in markdown fences and adding explanatory prose, which breaks a JSON.parse() call.

The reliable JSON prompt pattern

Three things make this work:

• The schema is defined upfront — field names, types, and null behaviour are all specified
• "No markdown code fences" — explicitly prevents the ```json wrapper Claude adds by default
• Format instruction is last — "Return only the JSON object. No explanation." closes the prompt

Defining the schema with an example

For complex or nested JSON, show Claude an example of the structure you want (few-shot from Chapter 4 applies here):

Return results in this exact JSON structure:
{
  "commands": [
    {
      "name": "string",
      "syntax": "string",
      "description": "string",
      "example": "string"
    }
  ]
}
No extra fields. No explanation. No markdown fences.

Markdown Tables

Markdown tables render cleanly in most documentation tools, GitHub, and note-taking apps. They're also easy for Claude to produce consistently.

CSV Output

CSV is slightly trickier because commas in values need quoting, and Claude may add a header row you don't want (or omit one you do). Be explicit:

HTML Fragments

For web projects, asking Claude to produce ready-to-use HTML saves a translation step. The key is specifying whether you want a full document or a fragment, and what CSS classes or structure to use.

Format Recipe Cards

Here are reliable prompt patterns for the most common structured output types:

Common Failure Modes

Structured Output in Pipelines

When Claude's output feeds directly into code, a few extra practices help:

• Parse defensively. Even with good prompting, occasionally Claude adds a preamble or trailing sentence. Strip whitespace and, for JSON, try parsing after removing any ``` fences as a fallback.
• Validate the schema. Use a library like Python's pydantic or JavaScript's zod to validate Claude's output against your expected shape before passing it downstream.
• Use the API's structured output feature. When calling Claude via the API, you can pass a JSON schema directly as a parameter — this enforces structure at the model level rather than relying purely on prompting. Covered in the Advanced API course.