From prompts to systems

The system prompt — a constitution

A system prompt is the persistent text that defines a bot's behaviour across every conversation it has. Where a one-off prompt asks the model to do one thing once, a system prompt tells the model who it is, what its purpose is, what it will and will not do — across thousands of users and millions of turns.

You write a system prompt when you:

• Create a custom GPT (in ChatGPT) or a Project (in Claude).
• Build an AI feature into a product you ship.
• Set up an internal assistant for your team or organisation.
• Configure an AI agent to handle a class of tasks reliably.

The five principles from Lesson 2 (Specificity, Context, Role, Format, Constraints) all apply to system prompts — but with more weight, because the prompt persists across many uses.

You are a meeting preparation assistant for a senior consultant at
[firm]. Your job is to take a meeting subject and a list of attendees,
and to produce a one-page prep document.

For each attendee, you produce:
- Name and role
- Two sentences on their likely concerns and motivations
- One concrete question the consultant could ask them

You also produce:
- Three potential meeting outcomes, ranked by likelihood
- Two questions the attendees may ask that the consultant should
  prepare answers for
- One thing the consultant should NOT do in this meeting

You do not:
- Speculate beyond publicly available information about specific people
- Use marketing language or hype
- Produce more than one page total
- Suggest specific deals, prices, or commitments

You always:
- Note your level of confidence in any background claim
- Suggest a structured agenda with time boxes
- Ask the consultant for clarification on anything ambiguous before
  producing the prep doc

If the consultant gives you incomplete information, ask for what
you need before proceeding. Do not invent context.

Notice: this looks like the long version of a one-off prompt. That is exactly what it is. The difference is that this text is invisible to whoever is using the bot afterwards — they just see an assistant that already knows how to do meeting prep. The discipline of writing the system prompt is the discipline of designing the bot's character.

Long-context prompts

Modern models can handle very large inputs — Claude can read entire books in a single context. This opens up a different kind of prompt: instead of asking a question and hoping the model knows the answer, you give it the document and ask questions about it.

The pattern.

You will be working with the document below. Your task is to
[specific job — summarise / extract / compare / fact-check / etc.].

Important reading instructions:
- Only use facts that appear in the document. Do not bring in
  outside knowledge.
- If a fact is implied but not stated, mark it as inferred.
- If a question cannot be answered from the document, say so
  explicitly — do not guess.

Cite the section or page number for every claim you make.

Document begins:
---
[paste the full document here, however long]
---
Document ends.

The specific task:
[your specific question, summarisation request, or analysis]

Three things to notice in this structure. First, the constraints come before the document — instructions placed near the start are usually weighted more heavily. Second, the explicit instruction to admit when something cannot be answered heads off the model's tendency to confabulate. Third, the citation requirement makes the model's reasoning checkable.

Long-context prompts are the primary technique for the kind of work covered in our Build a Market Research Bot, AI for Sport Analysts, and Reading an AI Paper Like a Researcher courses. Each of those is, at its heart, a long-context prompting exercise applied to a specific kind of document.

Prompts for agents

The frontier of prompt engineering in 2026 is prompts for agents — AI systems that take actions on your behalf rather than just chatting. Booking flights. Sending emails. Editing files. Calling APIs. Navigating websites.

Agent prompts have to do more than describe what you want as output. They have to describe what the agent should do, what tools it can use, what it must check before acting, when it should ask for confirmation, and what it must not do under any circumstances.

You are an agent that helps me manage my inbox. You have access to
my email account.

You will autonomously:
- Label and triage incoming messages
- Draft replies to routine messages (but NOT send them)
- Flag urgent messages for my attention
- Archive newsletters and notifications older than 14 days

You will NEVER, without explicit confirmation in the current
conversation:
- Send any message
- Delete any message
- Change inbox-level settings
- Reply to anyone outside the @[mycompany.com] domain
- Take action on any message marked "important" or "starred"

Before any action, you will state:
1. What action you propose to take
2. On which message(s)
3. The reason

If you are uncertain about any of those three, you stop and ask.

You log every action you take, with a one-line summary, to the
"Agent Log" folder.

Three principles are visible. Explicit allow-lists for what the agent can do. Explicit forbid-lists for what it must not do — including specific high-risk actions called out by name. A pause-and-confirm requirement before any action with consequences.

This style of prompt engineering — designing the agent's authority and constraints — is covered in much more depth in our Build Your First AI Agent course. It is one of the most-important emerging skills in AI.

How this course connects to the other seven

Every other course in our free library is, at some level, an application of prompt engineering to a specific domain. Knowing that lets you go back to those courses with sharper tools.

• Understanding AI — Lesson 1 of this course (what a prompt actually is) maps directly onto the foundational AI literacy in that course.
• Build Your First AI Agent — Lesson 5 of this course (agent prompts) is the technique that course is built on.
• Build a Market Research Bot — uses long-context prompting (Lesson 5 here) plus role-design (Lesson 2) plus the no-prediction constraint (Lesson 2's constraints principle).
• AI for Sport Analysts — same pattern as the market-research course, applied to sport data.
• AI Security Foundations — the OWASP "prompt injection" category in Lesson 2 of that course is, in part, an attack on the prompt-engineering layer.
• The EU AI Act for Non-Lawyers — the regulatory frame around what you can and cannot ask AI systems to do, including via prompting.
• Reading an AI Paper Like a Researcher — Lesson 3 of that course covers reading the method section, including the parts of academic papers that introduced techniques like chain-of-thought prompting (Lesson 3 of this course).

If you have taken or are planning to take any of those, you can now read them with prompt engineering as a connecting thread. The patterns recur because they work.

What we covered

Five lessons. What a prompt actually is. The five principles. The five patterns. How to iterate when prompts fail. How prompt engineering scales up to system prompts, long-context prompts, and agent prompts. You should now be able to write a prompt that consistently produces useful output, debug one that does not, and design the prompt-level behaviour of a custom bot or agent.

The skill compounds. The library compounds. The discipline you have just learned will be useful for every AI tool you encounter in the next decade.