One full lesson, and the prompts to go with it

01. Write the job down before you touch a builder

An agent is four things: a goal, the tools it may use, a loop, and a done-condition. Skip any one of them and you get a chatbot that costs money.

Write yours out in plain language first. For this build it is: answer one research question with sourced findings, using web search, in at most eight search steps, stopping when the brief is written or the step cap is hit.

02. Create the assistant and paste the instruction block

In whichever assistant builder you use, create a new assistant and put this in the persistent instructions field. It is deliberately blunt — instructions that read like a personality description produce agents that behave like one.

You are a research agent. You answer one question per run.

Process:
1. Break the question into at most 3 sub-questions.
2. Search for each. Maximum 8 searches total across the whole run.
3. Stop searching as soon as you can support a recommendation, even if that is after 2 searches.

Rules:
- Every factual claim carries a source link. No link, no claim.
- Never answer from memory. If search is unavailable, say so and stop.
- If sources disagree, report the disagreement. Do not resolve it yourself.
- If you cannot support an answer, return "No reliable answer found" plus the list of what you checked. This is a correct outcome, not a failure.

Output format, exactly:
RECOMMENDATION — one sentence
WHY — 3 bullets, each with a source link
TRADE-OFFS — 2 bullets
NOT VERIFIED — list, or "nothing"

03. Give it exactly one tool

Turn on web search. Turn everything else off. The most common cause of an agent behaving unpredictably is that it has five tools and no rule about which to reach for.

One tool per agent until it works. Add the second only when you can describe, in a sentence, the condition under which it should be chosen over the first.

04. Test it on a question you already know the answer to

This is the step that separates a demo from something you trust. Ask it something where you can personally verify the answer — a price you already know, a specification you have checked.

You are not testing whether it sounds good. You are testing three things: are the links real, does the recommendation match the evidence beneath it, and did it stop.

05. Break it on purpose

Now ask it something unanswerable — a fact that does not exist, or a question about a private company's internals.

A correct agent returns "No reliable answer found". If instead it produces a confident paragraph with plausible-looking sources, the refusal rule is not landing: move it higher in the instructions and state it twice. Agents weight the beginning and the end of an instruction block most heavily.

06. Give the output somewhere to land

Send the brief to a document, a note, or an email draft — anywhere a human reads it before it goes anywhere else.

That is the whole lesson in one line: the first version of any agent writes a draft. Automatic sending comes later, if ever. The six builds in the full course are this same shape with more moving parts.