Arkendia Classrooms
Lesson 7 of 21 · 15 min

Controlling the Output: Format, Templates, Tone and Length

A great answer in the wrong shape is still a problem: a wall of text when you needed a table, a casual tone for a formal letter, three pages when you wanted three lines. In this lesson you will learn to control exactly what comes back: its format, its template, its tone and its length, so that results are ready to use with minimal rework.

In chapter 2 you learned the five ingredients of a prompt. Here we zoom in on two of them, format and constraints, and turn them into a precise toolbox. Controlling the output is also the key to building reusable prompts for your personal work assistant: a prompt that always returns the same shape is far easier to reuse and to check.

Choosing the right format

The format should match what you will do next with the result. Ask yourself: where is this going? An email, a slide, a spreadsheet, a checklist, another prompt?

You need...Ask for...
A quick overviewA short bulleted list, 5 items maximum
A comparisonA table with named columns
A procedureA numbered list of steps
A document structureHeadings and subheadings (an outline)
Data to reuseCSV or JSON with defined fields
A message to sendSubject line, body and closing, with placeholders

Lists and tables

Lists make answers scannable; tables make comparisons clear. Always name the columns or items you want.

Compare these three laptop options for a student on a budget.
Format: a table with the columns Option, Price range, Strengths,
Weaknesses, Best for. Then add one sentence of recommendation.

Markdown

Many assistants format their answers with Markdown, a lightweight way to write headings, bold text, lists and tables with simple symbols (for example ## for a heading, **word** for bold, | to separate table columns). If you will paste the result into a document or a note-taking app that understands Markdown, ask for it explicitly. If you will paste it into an email, ask for plain text instead, to avoid stray symbols.

Structured data: JSON and CSV

If the output will be used by a spreadsheet or another program, ask for a strict structure. JSON is a common text format for structured data: fields and values in braces. CSV is a simple table with comma-separated values.

Extract the following information from the email below and answer
with valid JSON only, no commentary, using exactly these fields:

{
  "sender": "",
  "date": "",
  "request": "",
  "deadline": "",
  "urgent": true or false
}

If a field is not in the email, use null. Do not guess.

Email: [paste here]

Notice the three safeguards: "JSON only," exact field names, and a rule for missing data. Even so, always validate the result before feeding it to anything automatic: models sometimes break a format, add a sentence before the JSON, or invent a value to fill a field.

Templates: show the exact shape you want

The most reliable way to control structure is to provide a template with blanks to fill. The model follows it closely.

Write a meeting recap using exactly this template:

Meeting: [title]
Date: [date]
Decisions:
- [decision 1]
- [decision 2]
Actions (Who - What - By when):
- [owner] - [task] - [deadline]
Open questions:
- [question]

Use only the notes below. Write "none" for any empty section.

A template is just a few-shot example with blanks. It also makes outputs comparable from one use to the next, which helps when you evaluate your prompts later in this chapter.

Tone and style

Tone is easy to change once you name it precisely. Words such as "professional" or "friendly" work, but they are broad. Improve them with more specific pairs of ideas:

  • Register: formal, neutral, casual.
  • Attitude: warm, direct, enthusiastic, reassuring, firm but polite.
  • Reader: "for a customer who is upset," "for a colleague who knows the topic well."
  • Style rules: short sentences, active voice, no jargon, no exclamation marks, no clichés.

The best way to set a style is still by example: "Match the tone of the sample below." Provide a few sentences you wrote, and the model will imitate their rhythm and vocabulary.

Rewrite the message below for a customer who waited too long.
Tone: sincere, calm, solution-focused. No exaggerated apologies,
no corporate clichés. Maximum two short paragraphs.

Tip

Name what you want to avoid in the same breath as what you want. "Friendly but not chatty: no emojis, no exclamation marks" gives a much sharper target than "friendly" alone.

Controlling length

Length is one of the most common sources of disappointment. Models tend toward long, padded answers unless told otherwise. A few ways to control it:

  • Give a number and a unit: "maximum 120 words," "exactly three bullet points," "two sentences."
  • Limit structure: "one paragraph, no list."
  • Specify the level of detail: "headline summary only," "with one example per point."
  • Ask for progressive depth: "Give me a one-line answer first. I will ask for more detail if needed."

Be aware that models are not precise counters: a limit of 100 words may produce 90 or 120. Treat numbers as strong guidance and recount if it is critical. Asking for "no more than" is generally safer than asking for "exactly."

Giving priorities when constraints compete

When you stack several constraints (short, detailed, friendly, formal), the model has to make trade-offs. Help it by saying what matters most: "Priority order: 1) accuracy to the notes, 2) brevity, 3) tone." This is especially useful for longer, more complex prompts.

Separating instructions from material

When your prompt includes a text to process, show clearly where it begins and ends. Delimiters reduce confusion, particularly if the material itself contains sentences that look like instructions.

Summarize the text between the triple quotes in three bullet points.

"""
[text to summarize]
"""

Warning

A strict format does not make the content correct. A perfectly valid JSON can contain a wrong date, and a neat table can contain an invented figure. Format controls the shape; you still verify the substance, especially before the data flows into another tool.

Your turn: shape your answers

Exercise 1. Write a prompt that produces a weekly plan as a table with the columns Day, Main task, Time needed and Priority (high, medium, low), from a list of tasks that you provide. Limit the table to five days.

Exercise 2 (project). Create a template-based prompt for one task in your library (for example, an email reply or a meeting recap). Run it on two different inputs and check that the structure remains identical both times. Save it with the variable parts in square brackets.

See a solution

A possible answer for Exercise 1:

Role: You are a practical planning assistant.
Task: Organize the tasks below into a plan for Monday to Friday.
Format: a table with the columns Day, Main task, Time needed,
Priority (high / medium / low). One row per day, five rows only.
Constraints: no more than 6 hours of tasks per day; put high-priority
items first; do not add tasks I did not mention; if something does not
fit, list it under the table as "Not scheduled".

Tasks: [paste your list with rough durations]

For Exercise 2, if the structure changes between runs, tighten the template (exact labels, "use exactly this template") and add a rule for empty fields. This is a normal first improvement round.

Key takeaways

  • Choose the output format according to what you will do with it: list, table, steps, Markdown, JSON or CSV.
  • A template with blanks is the most reliable way to get a consistent structure; always specify what to do when information is missing.
  • Describe tone precisely (register, attitude, reader, style rules) and give a sample whenever you can.
  • Control length with numbers and structure limits, and state your priorities when constraints compete.
  • A correct format never guarantees correct content: validate before reuse.
Advertisement
Log in to save your progress. Log in