E1 · The reference work
30 minDo the task once yourself, the way you consider excellent. Save it as the reference and write down two mistakes you want to avoid.
Done when: there is a reference file made by you and a list of two mistakes to avoid.
Lesson 4 · 16 min
Anthropic, the company that makes Claude, recommends treating Claude like a brilliant but new employee who doesn't know your rules yet. Anyone who has hired people knows what that takes: a clear goal, the background of where it came from, and an example of work done well. In this lesson you put the three together into a role sheet and check your sheet right on the page.
Start the lesson
Goal: write an agent's role sheet with a role, a measurable goal, context, an example and a boundary.
Tap or click each step.
Scroll the page: each step below lights up its part of the diagram.
What needs to happen and how to know it happened. If you can't measure it, the agent can't check its own work and doesn't know when to stop (lesson 5).
Where this goal came from and why: who asked for it, who it's for, what has already been tried. Most of it comes from the base in lesson 3; in the sheet you just point to where it is.
"Do a great job" doesn't work: it doesn't know what great means to you. Do the task once yourself, done well. It's the best-spent time, and probably the last time you do this task.
The sheet brings the three steps together and adds two things every employee needs: what they must never do (boundary) and when you check the work. This is the sheet you hand to the agent.
A measurable goal, context and an example, plus a boundary and a check: the role sheet. You stop writing the work and start managing it.
A well-made sketch and a check once in a while: that is how you delegate to someone capable.
Cases told by Anthropic itself. Notice how short the requests are.
The creator of Claude Code (a programming agent) wanted the desktop app rewritten in Swift, another programming language.
The request came with its own test: run the old app, take a screenshot, compare the screens point by point with the new one, and don't stop until they match. The agent ran on its own for over two weeks.
The goal measures itself: the new screen is either the same as the old one or it isn't.
In another case, Claude joined a Slack conversation (the internal chat of many companies), pulled data from several tools and put together a single reply.
Nobody wrote the step-by-step: the role was to answer the team with data.
The context (access to the tools) was already there; the request was the team's question.
A researcher put sixteen Claude agents to work building a C compiler (the program that turns code written by people into a program the computer can run), with a simple order: keep going until it's perfect.
It took about two weeks of work.
"Perfect" there had a measure: the compiler passes the tests or it doesn't.
The requests are short, like the ones you would make to someone on your team. What they have is a way to measure the result and the context at hand.
This is what the role sheet forces you to write.
Short request + clear measure + available context = an agent that runs on its own. In the chat, the same sheet makes every conversation start already in the role; running on its own is something for the terminal and automations.
One line per field. Replace the text between < > with your own.
Role: <the job, in one sentence: you take care of ...> Goal: <what needs to happen> Measure: <how to know it is finished and good> Context: <where the rules, the lessons and the files are; why the task exists> Example: <where the reference work done by you is> Boundary: <what it must never do> Check: <when and how you look at the work>
From now on this is your role. Read the sheet below, the project rules and the example indicated. Before you start, tell me in three lines how you will check your own work against the Measure. <paste the sheet here>
Paste your sheet (or upload the .txt). The page checks the seven fields, rejects anything empty or vague and asks that the Measure include a number or a test along with what is being measured ("3 lines", "cites the date", "approved with no corrections"). The OK only checks the format; it doesn't guarantee the measure is a good one: you are the one who decides, in the practice below.
Use the "role candidate" you marked in lesson 1.
Do the task once yourself, the way you consider excellent. Save it as the reference and write down two mistakes you want to avoid.
Done when: there is a reference file made by you and a list of two mistakes to avoid.
Fill in role-sheet.txt with the seven fields and paste it into the checker above.
Done when: the checker answers "OK: sheet with all 7 fields filled in" and you wrote, in one sentence, how you check the Measure by looking at a deliverable.
Give the sheet to the agent (in the chat: paste it into the project instructions and open a new conversation; in the terminal: save it in the folder) and ask for the first run using the request from the template.
Done when: the agent described how it will check itself against the Measure and delivered a result that you compared with the reference.
Stop writing prompts: define the role.
The explanation appears after your answer.
Question 1 of 4
What are the three steps for delegating to an agent?skill
Question 2 of 4
Why do the task once yourself before delegating?
Question 3 of 4
In the app rewritten in Swift, what let the agent run on its own for two weeks?
Question 4 of 4
"Role: you are the best copywriter in the world." What's wrong with it?
Choose an answer. The explanation appears after you try.