L1Lesson 2Free

Your first agent: give it one job

After this, you'll have set up a single-purpose agent that perceives real input, decides what to do, and acts on it, all inside a Claude Project with custom instructions and a repeatable prompt.

Before you start

Complete What an AI agent actually is first. You need the perceive-decide-act frame to write useful instructions.

The idea

Your first agent is not a chatbot with a fancy title. It is a set of instructions that makes Claude do a specific job every time you run it. The secret: a Claude Project (a workspace inside claude.ai or the desktop app where you save custom instructions and files) turns a one-shot chat into a repeatable workflow.

An agent request card tries to enter three separate job lanes before one lane is chosen, with the golden dot on the single chosen lane only.

A single continuous line repeats into a neat stack of identical small shapes, the top one lifting away capped with a golden dot: the same small job done by hand over and over, every week. — The job you do by hand every week is exactly the job worth handing to an agent.

You write the instructions once, and every conversation inside that Project follows them automatically.

Here is the concrete job this lesson builds: a weekly email digest agent. You paste your recent emails into the Project, and the agent reads them (perceive), sorts them by urgency and topic (decide), then produces a clean summary with draft replies for anything that needs a response (act).

The whole thing takes about two minutes to run once it is set up, versus the twenty minutes of scanning and drafting you do manually. (Later in the track, you will wire real connectors so the agent reads your inbox directly. For now, paste is all you need.)

Agent StepsFollow the steps in order, then check You tested the agent with harder, messier input than.

You tested the agent withYou tested the agent with harder, messier input than the first run
You asked Claude toYou asked Claude to self-review its output against the Project instructions
Claude identified at leastClaude identified at least one drift and suggested an instruction fix
Open a new conversation inOpen a new conversation in your Project with fresh input you have not used before. If

The setup is three pieces:

Piece 1: The Project. In claude.ai, click Projects in the sidebar, then Create Project. Name it something specific ("Weekly Email Digest," not "My Agent").

In the desktop app, create a folder for this job instead. Either way, the Project is the container that holds your agent's instructions and remembers them across sessions.

Piece 2: The custom instructions. Inside the Project, open the Instructions panel (claude.ai) or the instructions file (desktop). These instructions tell Claude what to do every time a conversation opens in this Project.

For the email digest, the instructions say: "You are an email triage agent. When I paste emails, sort them into Urgent, Needs Reply, and FYI. For Needs Reply items, draft a short response in my voice."

One more line caps the length: "Keep the summary under 300 words."

Piece 3: The repeatable prompt. Each week, you open a new conversation in the Project and paste (or point to) your emails. Claude follows the instructions without you repeating them. That is the "agent" part: the instructions persist, the job repeats, the quality stays consistent because the rules do not change between runs.

The reason this works better than a single long prompt: a long prompt gets lost in the conversation. By the fourth reply, Claude has drifted from your original instructions.

Project instructions sit outside the conversation. They are re-read at the start of every new chat, so the agent stays on task even when the conversation gets long.

You can build this agent on any Claude surface (claude.ai Projects, the desktop app, or Google Antigravity). The lesson uses claude.ai Projects for the example because it requires no install and no code. The pattern transfers to the desktop app with zero changes.

Try it (12 min)

Watch out for

Fixing drift inside the conversation instead of in the Project instructions. The chat is temporary. The instructions persist. If you correct Claude in the chat, the correction disappears next session.
Writing instructions that are too vague ('be helpful' or 'summarize well'). Vague instructions produce vague output. Name the categories, the format, the word limit. Specific rules produce consistent agents.
Naming the Project something generic like 'Test' or 'Agent 1.' When you have five Projects, you will not remember which is which. Name it after the job it does.
Trying to build a multi-step agent that calls APIs or reads your inbox automatically. This lesson builds a paste-and-process agent. Automatic input comes in the next lesson when you add tools.

Paste this into Claude

Create your first agent Project. Follow these three steps:

STEP 1: CREATE THE PROJECT
Open claude.ai (or the desktop app). Create a new Project. Name it after the job, not "Test Agent." If you picked a task in the previous lesson, use that. If not, use "Weekly Email Digest."

STEP 2: WRITE THE INSTRUCTIONS
Open the Instructions panel in the Project and paste this (adapt the bracketed parts):

"You are my [task name] agent. Every time I start a conversation in this Project, follow these rules:

1. PERCEIVE: Read everything I paste carefully. Identify [what matters for your task].
2. DECIDE: Sort items into [your categories, e.g., Urgent / Needs Reply / FYI].
3. ACT: For each item that needs action, [describe the output, e.g., draft a reply in my voice, write a summary, flag the issue].

Keep your output under [word limit]. Use bullet points for the summary. Bold the most important item. Ask me before doing anything irreversible."

STEP 3: RUN IT
Open a new conversation inside the Project. Paste some real input (emails, notes, tasks, whatever fits your chosen job). See if the agent follows the perceive-decide-act loop you specified.

If it drifts from the instructions, do NOT fix it in the conversation. Go back to the Instructions panel and tighten the rule it broke. Then start a fresh conversation and test again. The fix lives in the instructions, not the chat.

What a good response looks like

## Weekly Email Digest, June 16

**Most important:** Client proposal deadline moved to Friday (was next Tuesday).

### Urgent
- **Proposal deadline change.** Acme Corp moved the submission window up by 4 days. Action needed today.

### Needs Reply
- **Team standup reschedule.** Priya asked to move Thursday's standup to 2pm. *Draft reply: "2pm works. I'll update the calendar invite."*
- **Vendor invoice query.** CloudHost billing asked about the May overage. *Draft reply: "The overage was approved in the May budget review. CC'ing finance for the PO number."*

### FYI
- Marketing newsletter went out (no action needed)
- Office Wi-Fi maintenance Saturday 6am-8am
- New hire orientation invite for July 1

**3 items need your attention. 2 draft replies ready for review above.**

What good looks like

You created a Project named after a real job (not 'Test' or 'My Agent')
The Project has custom instructions with perceive, decide, and act steps
You ran the agent at least once with real input and got structured output
When the output drifted, you fixed the instructions, not the conversation
The agent produced a result you would actually use (a summary you would read, a draft you would send, a list you would act on)

Go deeper (8 min)

Paste this into Claude

Now stress-test your agent. Open a new conversation in the same Project and paste a harder batch of input (more items, messier formatting, maybe a few that do not fit your categories cleanly).

After the agent responds, ask Claude this follow-up inside the conversation:

"Review what you just produced against the instructions for this Project. Did you follow every rule? Which rule did you follow least precisely? Suggest one change to the instructions that would prevent that drift next time."

Take Claude's suggestion, update the instructions, and run one more test. This self-review loop is how you tune an agent without guessing.

What good looks like

You tested the agent with harder, messier input than the first run
You asked Claude to self-review its output against the Project instructions
Claude identified at least one drift and suggested an instruction fix
You applied the fix to the Instructions panel (not inline in the chat) and re-tested

When this breaks

Breaks when the instructions try to do too many jobs at once. An agent that triages email AND writes blog posts AND tracks expenses will do all three poorly. One Project, one job.
Breaks when you skip the self-review step and just keep running the agent with the same instructions. The self-review is how the agent tells you what to fix. Without it, you are guessing at improvements.
Breaks when you expect the agent to remember previous conversations. Each new conversation in a Project starts fresh with the instructions. If you need memory across sessions, that is a later lesson (premium M4).

AI can help with this

Open Claude and say: 'I want to build an agent that [your job]. Write me the Project instructions using a perceive-decide-act structure. Keep it under 200 words.' Claude writes the instructions. You paste them into the Project and test. The whole setup takes under five minutes.

One narrow job card moves through input, action, output, and review with the golden dot on the finished artifact.

You can now

You have a working Claude Project that runs a specific job end to end

You can set up custom instructions that persist across conversations
You know to fix drift in the instructions panel, not in the chat
You ran the self-review loop and tightened at least one instruction

Key takeaways

An agent is instructions that persist. A Claude Project gives your instructions a home, so the job runs the same way every time without re-explaining yourself.

A Claude Project turns a one-shot chat into a repeatable agent. Custom instructions persist across every conversation inside the Project, so the agent stays on task.
Your first agent is three pieces: a Project (the container), custom instructions (the rules), and a repeatable prompt (the trigger). No code, no install.
When the agent drifts, fix the instructions, not the conversation. The conversation is temporary. The instructions are permanent.
The self-review loop (ask Claude to check its own output against the instructions) is the fastest way to tighten an agent without guessing.
One Project, one job. Multi-purpose Projects produce unfocused agents.

Go deeper

Was this helpful?

Up nextGive your agent tools

What an AI agent actually is (and what you'll build here)Thinking in Agentic Sessions

← Back to Build AI Agents

Create your first agent Project. Follow these three steps: STEP 1: CREATE THE PROJECT Open claude.ai (or the desktop app). Create a new Project. Name it after the job, not "Test Agent." If you picked a task in the previous lesson, use that. If not, use "Weekly Email Digest." STEP 2: WRITE THE INSTRUCTIONS Open the Instructions panel in the Project and paste this (adapt the bracketed parts): "You are my [task name] agent. Every time I start a conversation in this Project, follow these rules: 1. PERCEIVE: Read everything I paste carefully. Identify [what matters for your task]. 2. DECIDE: Sort items into [your categories, e.g., Urgent / Needs Reply / FYI]. 3. ACT: For each item that needs action, [describe the output, e.g., draft a reply in my voice, write a summary, flag the issue]. Keep your output under [word limit]. Use bullet points for the summary. Bold the most important item. Ask me before doing anything irreversible." STEP 3: RUN IT Open a new conversation inside the Project. Paste some real input (emails, notes, tasks, whatever fits your chosen job). See if the agent follows the perceive-decide-act loop you specified. If it drifts from the instructions, do NOT fix it in the conversation. Go back to the Instructions panel and tighten the rule it broke. Then start a fresh conversation and test again. The fix lives in the instructions, not the chat.

## Weekly Email Digest, June 16 **Most important:** Client proposal deadline moved to Friday (was next Tuesday). ### Urgent - **Proposal deadline change.** Acme Corp moved the submission window up by 4 days. Action needed today. ### Needs Reply - **Team standup reschedule.** Priya asked to move Thursday's standup to 2pm. *Draft reply: "2pm works. I'll update the calendar invite."* - **Vendor invoice query.** CloudHost billing asked about the May overage. *Draft reply: "The overage was approved in the May budget review. CC'ing finance for the PO number."* ### FYI - Marketing newsletter went out (no action needed) - Office Wi-Fi maintenance Saturday 6am-8am - New hire orientation invite for July 1 **3 items need your attention. 2 draft replies ready for review above.**

Now stress-test your agent. Open a new conversation in the same Project and paste a harder batch of input (more items, messier formatting, maybe a few that do not fit your categories cleanly). After the agent responds, ask Claude this follow-up inside the conversation: "Review what you just produced against the instructions for this Project. Did you follow every rule? Which rule did you follow least precisely? Suggest one change to the instructions that would prevent that drift next time." Take Claude's suggestion, update the instructions, and run one more test. This self-review loop is how you tune an agent without guessing.