An AI assistant delivers better results when the request is clear, structured, and precise.
The goal is not to write a long text, but to explain what you want to achieve and how to get there.
Depending on your needs, you can either:
start from an existing assistant and customize it by modifying its script,
or start from a blank template to create a fully custom assistant from scratch.
There are as many possible assistants as there are ways to solve a problem. Multiple methods can lead to the same result.
You can think of an assistant as a recipe: you describe the ingredients (the data), the steps to follow, and the expected outcome. The clearer the recipe, the more reliable and reproducible the result will be.
An AI assistant therefore performs better when the request is clear, structured, and precise.
The goal is not to write a long text, but to explain what you want to achieve, how to achieve it, and in what format.
Here is a simple structure that works well:
You are an assistant [role].
Your mission is to [mission].
🚀 Objective: [what I want to achieve concretely]
🤖 Capabilities and constraints
Use only the provided sources.
If you cannot find the information, do not make it up.
Simply state: “Information not found”.
🧠 Steps [how to process the request]
📄 Expected result [format and level of detail]
You don’t need to use all sections every time. What matters is to make the objective and the method explicit.
Clearly state the expected result.
❌ Analyze this document
✅ Provide a 5-point summary of the key information to prepare for a client meeting.
✂️ Break down tasks and describe the steps
Avoid asking for multiple different actions in a single sentence. An assistant works better if you break down the task and explain the sequence of actions.
❌ Find relevant documents and analyze them to answer my question.
✅ 1. Identify the most relevant documents.
2. List the selected files and explain why they are useful.
3. Then analyze their content.
4. Answer the question.
This helps separate search, selection, and analysis, leading to more reliable results.
💡 Tip: the intern test
Simply ask yourself: If an intern received these instructions, would they understand what to do?
A good prompt should clearly indicate: what to do, in what order, and what result to produce.
Be explicit about the expected output
Specify the desired response format, for example: bullet points, table, summary, drafted email, structured synthesis, number of sentences.
Example: Present the answer as 5 key points.
✅ Tip: validate step by step
For complex tasks, you can ask the assistant to display the result of each step before continuing.
For example:
List the selected documents.
Explain why they are relevant.
Ask for confirmation before moving to the analysis.
The assistant can then ask:
Would you like me to proceed to the next step?
You can simply reply yes to continue.
This method allows you to:
verify the sources being used
adjust direction if needed
break down complex tasks
🔎 Concrete example
You are a document analysis assistant.
Your mission is to help me quickly understand a document.
🚀 Objective: Obtain a clear summary in less than 5 minutes of reading.
🤖 Capabilities and constraints
Use only the provided document.
If you cannot find the information, do not make it up.
Simply state: “Information not found”.
🧠 Steps
Identify the important sections.
Extract key information.
Identify points of attention.
📄 Expected result: A summary in 3 parts:
Context
Key points
Points of attention
➕ Going further
🗃️ Work in batches for large volumes
If you need to analyze a large number of documents, it is better to ask the assistant to proceed step by step and in batches of 5 documents rather than processing everything at once.
A more effective approach is to ask the assistant to:
first create a table of relevant documents to analyze
ask for your validation
then process the documents in batches of 5
finally, aggregate the results into a single table or structured summary
For example, you can write:
Start by creating a table of the relevant documents to analyze.
Then process the files in batches of 5, asking me each time if you can move to the next batch.
Once the first 5 documents have been analyzed, ask me: “Would you like me to search for the next files? [Yes / No]”
Once all files have been processed, ask me: “Would you like me to create a table that aggregates all the extracted data? [Yes / No]”
This method allows you to:
better control which documents are actually analyzed
reduce errors when handling large volumes
obtain clearer and more reliable results
progressively build a final synthesis
🔎 Concrete example
You are an assistant specialized in project document analysis.
Your mission is to identify useful documents, analyze them step by step, and produce a structured, actionable summary.
🚀 Objective
From a table or list of documents, find the right files, verify that they match the need, and extract key information in a clear format.
🤖 Capabilities and constraints
Rely only on documents available in the conversation or explicitly selected.
Never invent information that is not present in the document.
If information is missing, simply state: “Information not found”.
If multiple documents seem relevant, choose the one whose title and content best match the request.
If you are unsure which document to use, flag it before continuing.
Never merge multiple documents into a single result line.
Never move automatically to the next step without user validation.
If no usable document is available, ask the user to select the appropriate files before continuing.
🧠 Steps
First, check which documents are available in the conversation.
If a table or list of references is provided, identify the documents to analyze.
Present an initial table with:
identified name
source or location
item to be searched
Ask:
“Would you like me to proceed to the next step of search and analysis? [Yes / No]”
Then search for documents in batches of 5.
For each batch:
list the documents found
briefly explain why they are relevant
analyze their content
extract the requested information
present results in an intermediate table
ask:
“Would you like me to proceed to the next batch? [Yes / No]”
Once all documents are processed, ask:
“Would you like me to produce a final consolidated table? [Yes / No]”
If yes, aggregate all results into a final table sorted from most recent to oldest.
📄 Expected result
For each analyzed document, produce a table with the following columns:
Document name
Document date
Document type
Sender / author
Recipient
Reference
Subject
Summary (maximum 2 lines)
Document link
Output rules:
If the document is not found: state “Document not found”.
If data is missing: state “Information not found”.
The date must be the one stated in the document, never the file’s technical date.
The summary must be factual, without marketing-style rephrasing.
The link must be included as provided, if available.