ابنِ أول agent لك من ملف واحد: مدرّب تلصقه في Claude أو ChatGPT أو GeminiBuild Your First Agent from One File: A Coach You Paste into Claude, ChatGPT or Gemini

Key terms

Agent

An AI assistant with one clear job, tools it can use, and a way to check its work before handing it to you.

Instructions

The fixed text the agent reads before every chat. It sets the job and the limits.

Markdown file

A plain text file ending in md. You write headings and normal lines, and any AI tool reads it easily.

Connector

A way to let the agent read from a tool you use, like email, calendar or files.

Test case

A real example from your work where you already know what a good result looks like.

The steps

1. 1

   Step 1

   Pick one task that repeats

   Your first agent should be small. Pick a task you do every day or every week and whose result you can judge: replying to inquiry emails, a meeting summary, a weekly report.

   Do not start with anything that touches money or big decisions.

   Tip: If you cannot explain the task in two lines, split it in two.

2. 2

   Step 2

   Paste the coach file into the chat

   Open Claude, ChatGPT or Gemini and paste the file below as it is. Then type "let's start".

   The coach asks one question at a time and writes no files until it understands the task. Then it writes four files and walks you through connecting tools.

   PromptThe coach file

   Copy

   You are a coach helping me build my first agent, and you write its files yourself. Do not write any file until the questions are done.
   
   Phase 1, interview (one question at a time):
   1. What is the one task that eats my time every day or every week?
   2. What goes in (files, emails, numbers) and what comes out?
   3. How do I know it was done right? Ask me for a real example of a good result.
   4. Who am I, and how do I like results delivered (language, tone, length)?
   5. What must the agent never do, and what must it ask me before it acts?
   
   Phase 2, split: break the task into numbered steps and mark each one "agent" or "me".
   
   Phase 3, files: write four markdown files, each in its own block with its exact name:
   IDENTITY.md: who the agent is, its one job, and its tone.
   USER.md: who it works for, who I am, and my preferences.
   MEMORY.md: the facts it must remember, each with a date. This file is updated as it works: whenever it learns something new, write me the new lines and I paste them into the file.
   AGENTS.md: its steps in order, its rules, and what it must ask me before it acts.
   Do not invent anything I did not tell you. Where something is missing, write [missing] and ask me about it.
   
   Phase 4, tools: ask me which tool this task needs (email, calendar, files). Then walk me step by step through connecting it in the app I am using now:
   Claude: Customize, then Connectors.
   ChatGPT: chatgpt.com/plugins, then type @ in the message box and pick the tool.
   Gemini: the Connected Apps setting.
   The rule: start with one tool and read only access. Do not suggest a second tool until the first works. And add a rule to AGENTS.md saying the agent never writes, sends or deletes.
   
   Phase 5, test: give me a test case from what I told you and how to judge the result.
   
   Start when I type "let's start".

3. 3

   Step 3

   Answer with real examples

   When it asks what a good result looks like, paste a real one from your work instead of describing it. One example teaches the agent more than a paragraph.

   Remove names and sensitive numbers before you paste.

4. 4

   Step 4

   Understand the four files

   The coach does not write your instructions as one block of text. It writes four small files, and each has one job.

   IDENTITY.md: who the agent is. Its one job and its tone. Think of an ID card.

   USER.md: who it works for. You, your work and your preferences. It knows you without you repeating yourself.

   MEMORY.md: what it must remember. Facts about your work and past decisions. It updates over time: when the agent learns something new, it gives you the lines and you add them to the file.

   AGENTS.md: the work manual. Its steps in order, its rules, and what it must ask you before it acts.

   Tip: When you change something later, you usually edit one file only.

5. 5

   Step 5

   Put the files where they live

   When the coach finishes, copy each block into a text file with the exact same name: IDENTITY.md, USER.md, MEMORY.md and AGENTS.md.

   In Claude: create a Project, add the four files to the project files, and paste the line below into the project instructions.

   In ChatGPT: create a GPT or a Project, upload the four files, and paste the same line into the instructions.

   In Gemini: create a Gem, upload the four files, and paste the same line into the instructions.

   Now every new chat starts with the agent knowing who it is, who you are and its job.

   support.claude.com

   

   PromptInstructions line

   Copy

   Read IDENTITY.md, USER.md, MEMORY.md and AGENTS.md before every chat and follow them. If anything is missing, ask me before you act.

   Tip: Name the project after the task, not "agent 1".

6. 6

   Step 6

   Connect one tool, read only

   Now connect the agent to a tool. Start with one tool and read only access: it reads, it never writes or sends.

   The coach walks you through it in the chat. These are the places:

   Claude: Customize, then Connectors.

   ChatGPT: chatgpt.com/plugins, then type @ in the message box and pick the tool.

   Gemini: the Connected Apps setting.

   When you connect the tool, choose read permissions if it gives you the option. If it does not, rely on the AGENTS.md rule that stops the agent from writing, sending and deleting.

   Try the tool on today's task. When it runs with no problems, add a second tool.

   support.claude.com

   

   Tip: One tool that works beats five tools connected wrong.

7. 7

   Step 7

   Test it on today's task

   Open a new chat inside the project and paste today's real case. Compare the result with the good example you gave.

   Do not judge from the first reply alone. Find exactly where it went wrong.

8. 8

   Step 8

   Fix three things after the first run

   Most problems come from three places: wrong format, missing information, or it did something you did not ask for.

   Go back to the coach chat, tell it what went wrong, and let it rewrite the right file for you instead of editing it yourself. It is usually AGENTS.md.

   PromptFix after the test

   Copy

   I tested the agent on this case: [case]
   This is the result: [result]
   The problem: [wrong format, missing information, or it did something I did not ask for]
   Update the right file so it does not happen again, and tell me which file you changed and exactly what you changed.

9. 9

   Step 9

   Know where to stop

   Never give the agent anything that pays money, sends in your name without your review, or deletes files. Let it prepare, and you press send.

   Connect one tool first, read only, then add. Do not open every tool at once.