ابنِ مساعد AI لشركتك على خمس مراحل: من موديل جاهز إلى agent يشتغل بصلاحيات الموظفBuild Your Company's AI Assistant in Five Stages: From a Ready Model to an Agent That Works With the Employee's Own Permissions

Key terms

API

How your software talks to a ready model like Claude. You send the question and get the answer back.

System prompt

A fixed text the model reads before every conversation. It holds the assistant's name, tone and limits.

RAG

The assistant searches your company files and pulls the right passage before it answers, so the answer comes from the source, not from general memory.

Evals

A set of questions with known correct answers that you run on the assistant before every change, to make sure nothing broke.

Agent

An assistant that uses tools and takes steps on its own, like opening a ticket or filing a request, not only writing an answer.

The steps

1. 1

   Step 1

   Stage one: connect a ready model through the API and give it one job

   Start with the simplest version: a ready model like Claude through the API, and a small interface where staff type a question.

   Do not try to cover everything. Pick one clear job, for example HR questions only. You will notice its answers are generic. That is expected, and it tells you what the next stage is.

   PromptPick the job before you build

   Copy

   I am building an internal AI assistant for my company [company name]. We work in [industry] and we are about [number of employees] people.
   Help me pick one job for the assistant to start with. Suggest 3 options, and for each one: who uses it, the 5 most likely questions, and which files it will need later.

   Tip: One clear job is far easier to test than an assistant that answers everything.

2. 2

   Step 2

   Stage two: write the system prompt and add conversation memory

   Now give it an identity: its name, its tone, the language it answers in, and what it does when it does not know.

   Add conversation memory so it continues from what the employee said instead of starting over every message. It still does not know your files, so the instructions must tell it to say so instead of making things up.

   platform.claude.com

   

   PromptDraft the system prompt

   Copy

   Write a system prompt for an internal assistant called [assistant name] at [company name].
   Tone: [formal / friendly]. Language: [Arabic / English / match the question].
   Its job: [the job from stage one].
   Fixed rules: never invent numbers or policies; if it does not have the information it says so and points the employee to [team]; never ask for personal data it does not need.

   Tip: One line that says «if you do not know, say so» saves many mistakes before the files are connected.

3. 3

   Step 3

   Stage three: connect it to company files (RAG)

   Collect the files that answer the job's questions: policies, forms, FAQs. Split them into small passages and let the assistant search them and pull the right passage before it answers.

   Ask it to cite the source in every answer: the file name and section. Staff can check it themselves, and you know where each answer came from.

   PromptPrepare the files before connecting

   Copy

   Here is the list of [department] files we have: [file names].
   Rank them by importance for an assistant that answers questions like: [3 expected questions].
   Tell me which files are outdated, duplicated or contradict each other, and what information is missing that we should write before connecting them.

   Tip: An outdated file in the source gives an outdated answer with full confidence. Clean the files first.

4. 4

   Step 4

   Stage four: log every question and answer, and write your evals

   Before anything else, you need to know when it is wrong and why. Log every question, every passage it retrieved, and every answer.

   Then write a list of questions with known correct answers and run it on the assistant after every change to the instructions, the files or the model. Any change that breaks an answer that used to be right does not ship.

   platform.claude.com

   

   PromptWrite your first 20 evals

   Copy

   From these files: [paste the passages or file names]
   Write 20 test questions for a [job] assistant. For each one: the exact correct answer, the file and section it comes from, and the type (direct, needs two files, or a question it must decline).
   Make 4 of them about things that are not in the files, so we can check it says it does not know.

   Tip: The questions it must decline matter as much as the ones it must answer.

5. 5

   Step 5

   Stage five: let it act, with the employee's own permissions

   Only now, with logs and evals in place, give it tools: open a ticket, file a leave request, book a meeting room.

   Every tool runs with the permissions of the employee talking to it, never higher. Sensitive steps like sending or approving ask the employee to confirm first. Add tests for the tools themselves to the same eval list.

   PromptTools and permissions plan

   Copy

   I want to give the [job] assistant these tools: [list of tools, like opening an IT ticket or a leave request].
   For each tool tell me: what data it needs, which employee permission is enough, whether it needs a confirmation before it runs, and the worst thing that could happen if it is misused and how we prevent it.
   Then write 5 tests per tool, including one where an employee asks for something they have no permission for.

   Tip: If the employee cannot open a file themselves, the assistant should not open it for them.

6. 6

   Step 6

   Do not jump to stage five before stage four

   The most common mistake: the team builds the tools first because that is the fun part, and testing gets pushed back.

   Anthropic's own advice in Building effective agents: start with simple prompts, optimize them with comprehensive evaluation, and add multi step agentic systems only when simpler solutions fall short. Use the prompt below as a quick review before any new stage.

   anthropic.com

   

   PromptReview before the next stage

   Copy

   Here is where our internal assistant is now: [current stage, what is connected, how many evals we have, and the pass rate].
   We want to move to [next stage].
   Review our plan honestly: is the current stage complete? What is missing? And what is the smallest version of the next stage we can try with a small group of staff?

Ready prompts to copy

We have an internal AI assistant. Here is what it does now: [description]. Here is the problem staff complain about most: [problem].
Tell me which of the five stages it is at: (1) model + API, (2) instructions and memory, (3) company files, (4) evals and monitoring, (5) tools with the employee's permissions. And the first step that fixes the problem.