Make AI Admit It Does Not Know: Grounding Rules and the Instructions File

The Real Goal: An AI That Admits the Limits of What It Knows

By default, an AI model is built to always produce an answer. Ask it something it does not know, and it will not stop and admit the gap - it will generate the most plausible-sounding response it can, presented with full confidence. That is hallucination, and it is the single most dangerous failure mode for anyone using AI for real work, because a confident wrong answer is far more harmful than an honest I-do-not-know. The goal of this guide is to reach a specific, advanced state: an AI that, when it genuinely does not have the answer, tells you so plainly instead of inventing one. This does not happen by choosing a smarter model. Even the best model will fabricate when you give it a vague question and no boundaries. It happens by changing how you use the AI - by constraining it so honesty becomes the only path it is allowed to take. That constraint is called grounding, and the rest of this guide is how to build it.

Why You Must Use Agent Mode, Not the Chat Window

Grounding is impossible inside the Chat window, for a simple structural reason: Chat has no access to your real files. It answers from the model's memory, so when a gap appears, it has nothing to fall back on except fabrication. The agent environment is fundamentally different. It opens a folder on your computer and works inside it - reading your actual files, your real documents, your verified data. This mode exists across all three major companies under different names: Cowork for Claude, Codex for the ChatGPT ecosystem, and Antigravity for the Gemini ecosystem. The same shift, the same principle, three brands. This is not a niche developer feature anymore. Globally, serious AI users are moving away from the Chat window entirely and toward the agent environment connected to their files. Arabic-language content has largely not caught up to this shift yet - which means the people who adopt it now are early, and early is an advantage. The Chat window is becoming the past. The grounded agent environment is becoming the standard.

The Instructions File: The Lever Nobody Uses

When you open a conversation inside a folder, that folder can hold a special instructions file. In Claude it is CLAUDE.md; the direct equivalents are AGENTS.md for Codex and GEMINI.md for the Gemini tools. Here is the critical fact most people never learn: the companies that built these tools wrote a hard rule into the model itself - it is not permitted to answer your question before it has read that instructions file. The model reads it first, every single session, automatically. This makes the instructions file the single most powerful lever you have over the AI's behavior, because whatever you write there shapes every answer that follows. And yet almost everyone leaves it empty, or fills it with vague notes. That is a wasted opportunity of enormous size. The instructions file is where you set the fundamentals of how the AI is allowed to work for you. Treating it as an afterthought is the difference between a casual user and someone who has genuinely configured their AI.

Writing Grounding Rules That Actually Hold

The instructions file only delivers if you fill it with strong grounding rules. A few rules carry most of the weight. The source-of-truth rule: the AI answers only from the files in this folder, and those files override its general knowledge. The honesty rule, the most important one: if the answer is not in the files, the AI must say plainly that it does not know - and you state that an honest I-do-not-know is a correct, acceptable answer, so it has explicit permission to stop. The no-fabrication rule: never invent numbers, dates, names, citations, or sources; a missing detail is reported as missing, not estimated. The citation rule: every factual claim names the file it came from, and if no file can be cited, the claim is not made. The knowledge-separation rule: if general knowledge is ever used, it must be clearly labeled as such so you always know what is grounded and what is not. Write these rules short and absolute - vague wording gets ignored by the model. Then test the setup: ask something you know your files do not cover, and confirm the AI admits it does not know. If it fabricates instead, your rules are too soft. Tighten them and test again. The instructions file is a living document - every failure you notice is a rule you have not written yet.

خلّي الذكاء الاصطناعي يقلّك «ما بعرف»: قواعد التأريض وملف التعليمات

هاد الدليل مو هدفه الانتشار أبداً. هدفه إنه يعلّمك مهارة وحدة محدّدة ومتقدّمة: كيف توصل للمرحلة اللي لما تسأل فيها الذكاء الاصطناعي عن شغلة ما بيعرفها فعلاً، يقلّك بصراحة «أنا ما بعرف»، بدل ما يخترعلك جواب واثق وغلط.

هالسلوك ما بيصير بالصدفة. لحتى توصل إله، لازم تجبر الذكاء الاصطناعي يجاوبك من مواد موجودة فعلاً على كمبيوترك بس. ولحتى تعمل هيك، لازم توقف تستخدمه ضمن واجهة الـ Chat، وتبلش تستخدمه ضمن بيئة الوكيل: Cowork لـ Claude، وCodex لمنظومة ChatGPT، وAntigravity لمنظومة Gemini. شباك الـ Chat عم يصير من الماضي. بيئة الوكيل، الموصولة بملفاتك الحقيقية، عم تصير المعيار بكل مكان، حتى لو المحتوى العربي لسا ما لحق هالتحوّل.

هون الفكرة الأساسية. لما تفتح محادثة جوّا مجلد على كمبيوترك، هالمجلد فيه ملف تعليمات: CLAUDE.md لـ Claude، وإله مكافئات مباشرة متل AGENTS.md وGEMINI.md لباقي الأدوات. الشركات اللي ورا هالأدوات حطّت قاعدة صارمة جوّا الموديل نفسه: ممنوع عليه يجاوب سؤالك قبل ما يقرأ ملف التعليمات.

هالقاعدة هي الرافعة. معظم الناس بيتركوا ملف التعليمات فاضي أو بيتجاهلوه كلياً. التحرّك الاحترافي هو إنك تعبّيه بقواعد تأريض صارمة (Grounding Rules): القواعد اللي ما إلها نقاش وبتحكم كيف مسموح للذكاء الاصطناعي يجاوبك. قواعد بتجبره يستند بس على ملفاتك الموثّقة، يستشهد بمصدر كل معلومة من وين جابها، ويقلّك «ما بعرف» لما الجواب ما يكون قدّامه.

هالممارسة الوحيدة هي الفرق بين ذكاء اصطناعي بيضلّلك بثقة، وذكاء اصطناعي فعلاً بتقدر تثق فيه.

تابع حسابي على الانستغرام:

رابط الانضمام للدورة: https://halaqa.app/enrollment?course=start-with-ai

الخطوات

Prompt

# GROUNDING RULES: THE INSTRUCTIONS FILE TEMPLATE
# File name by tool:
#   Claude Code / Cowork  → CLAUDE.md
#   Codex (ChatGPT)       → AGENTS.md
#   Gemini CLI / Antigravity → GEMINI.md
# Place it in the root of your project folder. The model reads it FIRST.

# ════════════════════════════════════════
# PASTE THIS INTO YOUR INSTRUCTIONS FILE
# ════════════════════════════════════════
"""
# GROUNDING RULES: READ BEFORE ANSWERING ANYTHING

## Source of Truth
- Answer ONLY from files inside this folder and its subfolders.
- These files are the single source of truth. Do not override them
  with general knowledge.

## The Honesty Rule (most important)
- If the answer is NOT found in the provided files, say exactly:
  I do not know, this is not in the provided files.
- Never guess. Never invent. A clear I-do-not-know is a correct answer.
- Do not produce confident answers to fill a gap.

## No Fabrication
- Never invent facts, numbers, dates, names, citations, or sources.
- If a detail is missing, state that it is missing. Do not estimate it.

## Citation Rule
- For every factual claim, name the file (and section) it came from.
- Format: [source: filename.md > section]
- If you cannot cite a file for a claim, do not make the claim.

## Separating Knowledge Types
- If I explicitly ask for general knowledge, you may use it,
  but you MUST label it clearly: [general knowledge: not from your files].
- By default, assume I only want answers grounded in my files.

## When Files Conflict
- If two files disagree, do not pick one silently.
- Show both, name both files, and ask me which is correct.

## Scope Discipline
- If I ask about something outside the provided files, tell me it is
  outside the grounded scope before answering anything.
"""

# ─── HOW TO USE IT ───
# 1. Create the instructions file in your project folder.
# 2. Paste the grounding rules above. Adjust to your needs.
# 3. Add your actual reference files to the folder (docs, data, notes).
# 4. Open the folder in Cowork / Codex / Antigravity, NOT the Chat app.
# 5. Ask your questions. The model reads the rules first, every session.

# ─── HOW TO TEST IT WORKS ───
# Ask a question you KNOW is not covered by your files.
# A correctly grounded AI will tell you it does not know.
# A failing setup will invent an answer. Fix your rules and retry.

# ─── WHY CHAT MODE CANNOT DO THIS ───
# Chat mode  → no folder, no files, no instructions file → it answers from
#              memory and fills gaps with plausible-sounding fabrication.
# Agent mode → reads your files + your grounding rules first → it answers
#              from verified material, and admits the gaps honestly.

# ─── PRO TIPS ───
# Keep the rules short and absolute. Vague rules get ignored
# Test the I-do-not-know behavior before trusting the setup on real work
# One instructions file per project: tailor the rules to that project
# Update the rules when you notice a failure pattern: it is a living file