CourseLevel 4Guide 17
17🤖 Level 4 — Build Your AI Team

Agent Instructions That Actually Work

20 min readIntermediate

What you will learn

  • The five parts every solid agent instruction set needs, and how to write each one so the agent behaves consistently instead of improvising.

Difficulty: Intermediate · Estimated time: 20 min

What You'll Learn

The five parts every solid agent instruction set needs, and how to write each one so the agent behaves consistently instead of improvising.

Why This Matters

A rough instruction set like the one from the last guide works until it doesn't — the agent drifts, guesses at formatting, or wanders into territory you didn't want it in. The fix isn't more instructions, it's better-structured ones.

The Concept

Good agent instructions have five parts, in this order:

  • Role — who it is.
  • Objective — what it's for.
  • Constraints — what it must never do.
  • Inputs — what it needs from you every time.
  • Output spec — exactly what the result should look like.
  • Skip the constraints and the agent will happily do things you never wanted (sending instead of drafting, inventing figures it doesn't have). Skip the output spec and every result comes back shaped differently. The structure does the consistency work so you don't have to re-check it every time.

    Step-by-Step

  • Role — one or two sentences. Name the specialism, not a generic helper.
  • Objective — what success looks like for this agent, in concrete terms ("produces a report my manager can forward without edits," not "helps with reports").
  • Constraints — a short, explicit list of what it must never do. Include anything that would be a real problem if it went wrong: never send on my behalf, never invent numbers, never contact a client directly, always flag anything it's unsure about.
  • Inputs — list exactly what you'll hand it each time, and in what form (pasted text, an attached file, a link).
  • Output spec — format, length, tone, and any labels or headings it must always include. The more exact this is, the less you'll need to edit afterwards.
  • Read the finished instructions back as if you'd never seen the task before — if anything's ambiguous to you, it'll be ambiguous to the agent too.
  • Example

    Weekly report writer, fully specified:

    Role: You are my weekly status report writer. You turn my rough bullet-point notes into a polished update for my manager.

    Objective: Produce a report my manager can read in under two minutes and that clearly separates what's done, what's in progress, and what's blocked.

    Constraints: Never add achievements or numbers I haven't given you. Never soften or hide a blocker — flag it plainly. Never exceed 300 words.

    Inputs: My raw bullet notes for the week, pasted as plain text.

    Output spec: Three headings — Done, In Progress, Blocked — each with up to four short bullet points. Plain, factual tone, no filler phrases.

    Try This

    Take the one-sentence job and role you wrote in Guide 16's "Try This" and expand them into all five parts now. Write the constraints section first — it's the one people skip, and it's the one that saves you the most trouble later.

    Copy This

    ROLE:
    You are [SPECIFIC ROLE]. You [WHAT IT DOES DAY TO DAY].
    
    OBJECTIVE:
    Success looks like: [CONCRETE, CHECKABLE DEFINITION OF A GOOD RESULT].
    
    CONSTRAINTS — NEVER:
    - [THING IT MUST NEVER DO, e.g. send anything on my behalf]
    - [THING IT MUST NEVER DO, e.g. invent facts or figures not supplied]
    - [THING IT MUST NEVER DO, e.g. exceed X words/length]
    
    INPUTS YOU WILL RECEIVE EACH TIME:
    - [INPUT 1, with format — e.g. pasted text, attached file]
    - [INPUT 2]
    
    OUTPUT SPEC:
    - Format: [HEADINGS / SECTIONS REQUIRED]
    - Length: [LIMIT]
    - Tone: [DESCRIPTION]
    - Always include: [ANY REQUIRED LABEL, FLAG OR DISCLAIMER]

    Common Mistakes

  • Writing constraints as vague wishes ("be careful with numbers") instead of hard rules.
  • Leaving the output spec loose and then being annoyed when formatting varies.
  • Bundling two different jobs into one role instead of splitting them into two agents.
  • Never revisiting the instructions after the first version — treat them as a draft, not a finished product.
  • Level-Up Tip

    When an output comes back wrong, don't just fix that one result — go back and add the missing rule to the instructions, so the same mistake can't recur next time.

    Action Checklist

    ○Role written as a specific specialism, not a generic helper
    ○Objective stated as a concrete, checkable outcome
    ○At least three hard constraints listed
    ○Inputs and output spec both written out in full
    ○Full instruction set re-read for ambiguity

    Next Guide

    18 — Give Your Agents the Right Context, where you'll build the background material that makes an agent sound like it actually knows your world.


    Premium Content

    Levels 2–6 require a mygyaan membership to access the full guides.

    ₹1,000/month — full mygyaan.ai access (all modules)

    ₹3,000 one-time — this course only

    Free, forever

    Sign up free for mygyaan.ai

    One book, one learning story, one quote, one video — delivered every day at 7 am. Unlimited books and turnaround stories, no payment, ever.

    Free forever. By signing up you agree to our Terms and Privacy Policy.

    No password. No spam. Unsubscribe any time.