Output Formatting: Get AI Answers in the Exact Structure You Need

You'll learn: how to specify the exact output format — tables, strict templates, or JSON — so AI answers are consistent, scannable, and ready to use or automate. You'll walk away with: the Output Formatting Toolkit — the format menu + the JSON & template patterns. Level: Beginner → Intermediate · Prereq: Turn a weak prompt into an expert one

Every example below is a real, unedited Claude (Sonnet 4.6) response from June 2026. Same data — three coffees — formatted four different ways.

1. The Problem

You ask AI for something useful and get back a wall of text you then have to chop into a table, a list, or a tidy template by hand. The content was fine; the shape was wrong.

Here's the fix most people skip: you can specify the exact format you want, and the AI will deliver it — a comparison table, a strict template repeated for every item, even clean JSON you can feed into a spreadsheet or your website. Same content, the shape you actually need.

Our data for all four runs:

Our three coffees:
- Ethiopian Yirgacheffe — light roast — blueberry, citrus, tea-like — $19
- Colombian Huila — medium roast — caramel, red apple, cocoa — $17
- Sumatra Mandheling — dark roast — dark chocolate, cedar, earthy — $18

2. Beginner Example — What "No Format" Gets You

[the data] … Tell me about them so a customer can choose.

What happened: modern Claude didn't just dump a paragraph — it produced nicely designed cards (it even added a "most popular" badge) plus a prose decision guide: "If they like bright, tea-style drinks → the Yirgacheffe… if they want something bold → the Sumatra."

Impressive! But here's the catch: the AI chose that structure, not you. That's a problem for three reasons:

• It's unpredictable — ask again tomorrow and you might get a totally different shape.
• It may not fit where you need it (a spreadsheet? a database? a uniform set of cards?).
• You can't reuse it reliably across many items.

Leaving the format unspecified means handing the AI the steering wheel. Let's take it back.

🔑 No format = the AI's choice. It might look great — but "looks great once" isn't the same as "the exact shape I need, every time."

3. Intermediate Example — Ask for a Table

[the data] … Present them as a comparison table with these columns:
Coffee, Roast, Tasting notes, Best for, Price.

What happened: a clean, scannable comparison table — exactly those columns. It even filled the "Best for" column intelligently (Yirgacheffe → "Pour-over lovers… bright, fruit-forward cups"), inferring it from the flavor profiles.

Naming the columns is the whole trick. You decided the shape; the AI filled it.

🔑 Tables: say "as a table with columns: X, Y, Z." Instant, scannable comparisons.

4. Advanced Example — A Strict, Repeatable Template

When you need every item in an identical shape (cards, listings, entries), give a template and say "use EXACTLY this."

[the data] … Format each coffee as a product card using EXACTLY this template:

**[Name]** — [Roast]
Tastes like: [notes]
Best brewed: [method]
Price: [price]
In one line: [a short Brew Lab-style one-liner]

What happened: all three came out in the same structure, with on-brand one-liners:

"Like a morning in Addis Ababa — floral, bright, and gone too soon." "The crowd-pleaser that never had to try — smooth, warm, and always right." "For those who want their coffee to mean it."

Now it's consistent and reusable — paste in 50 products and every card matches.

🔑 Strict templates: "use EXACTLY this template" → identical structure across every item.

5. Power Move — Ask for JSON

This is the one that unlocks automation. Ask for JSON and you get machine-readable data you can drop straight into a website, spreadsheet, or workflow:

[the data] … Return the data as a JSON array. Each object must have these keys:
name, roast, tasting_notes (array of strings), best_brew_method, price_usd (number),
tagline. Return only valid JSON — no extra text.

Real output:

[
  {
    "name": "Ethiopian Yirgacheffe",
    "roast": "light",
    "tasting_notes": ["blueberry", "citrus", "tea-like"],
    "best_brew_method": "Pour over",
    "price_usd": 19,
    "tagline": "Bright and delicate with a fruit-forward finish"
  }
]

Two things made this work:

• Name the keys and types ("price_usd (number)", "tasting_notes (array of strings)") so the structure is exactly what your system expects.
• "Return only valid JSON — no extra text" stops the AI from wrapping it in chatty commentary.

This is the bridge to real automation — feed AI output into other tools without copy-pasting.

🔑 JSON: specify keys + types and "return only valid JSON." Now the output is data, not just text.

6. ⚠️ Watch the Inferred Fields

Notice the table had a "Best for" column and the JSON had best_brew_method — neither was in our original data. The AI filled them by inference. They're reasonable, but they're guesses, not facts you gave it.

This is the same gap-filling behavior from Context Prompting: when your format asks for a field your data doesn't contain, the AI invents a plausible value. Verify inferred fields before you publish or pipe them into a system.

7. Common Mistakes

8. Templates (Your Take-Home)

Table: Present this as a table with columns: [A], [B], [C].

Strict template:

Format each [item] using EXACTLY this template:
[your template with [placeholders]]

JSON:

Return the result as a JSON array. Each object must have these keys:
[key (type), key (type), ...]. Return only valid JSON — no extra text.

📥 Download the Output Formatting Toolkit (free) — the full format menu (prose / list / table / template / JSON / markdown) with when-to-use. (Email opt-in.)

9. Your Challenge

Do this now: take any list you have (tasks, products, contacts). Ask the AI to return it three ways — a table, a strict template, and JSON with named keys.

You did it right if: the JSON is clean enough to paste into a spreadsheet — and you spotted any field the AI inferred rather than pulled from your data.

Keep going: ← Pillar: Upgrade a Prompt · Siblings: Role Prompting · Context Prompting · Few-Shot Examples · Start the Starter Course →