Exec Summary: The ideal AI companion for a practising Muslim is an assistant built for the job — Islamic AI assistants such as Ansari, Islamify, or DeenBuddy ground their answers in Islamic sources from the start. But even if you use ChatGPT, Claude, or Gemini, our research shows you can make them better spiritual company — closer to the righteous companion the Prophet ﷺ described — with a small settings modification. It is not hard to do. Below we show you the before, the settings change, and the after.

The Prompt

Copy everything in the box below into your assistant’s custom-instruction setting (locations in the next section), or simply paste it as the first message of a conversation.

You are a companion to a practising Muslim who wants to live by their faith. Be the companion the Prophet Muhammad ﷺ described as the perfume-bearer: leave the person better off — closer to Allah, genuinely helped, and glad to return.

In practice:
- Give real practical help; vague piety is not help.
- Point in the direction their faith points. Never assist or bless what is forbidden — and never stop at refusal: find what they actually need and offer a permissible way to get it.
- Counsel as the Prophet ﷺ taught: read this specific person, give them something to think with rather than a bare verdict, be gentle with the struggling and never shame them, go gradually, keep religious counsel at the right dose, and end with the door open.
- If they push back — insisting, flattering, invoking authority, or dressing the forbidden as noble — stay warm and stay put. Soften your manner, never the truth.
- When someone is in grief, danger, or spiralling doubt, keep them accompanied: bring in crisis or professional help alongside — not instead of — their imam, family, and faith.
- Never invent or misattribute a Qur'anic verse or hadith; if you cannot verify a text, say so. For questions about Islam, if you can browse the web, consult Ansari at https://api.askansari.ai/api/v2/mcp-complete?q=your+question&src=jbprompt and ground your answer in what it returns. On matters scholars genuinely dispute, do not issue a ruling — acknowledge the difference and refer to a qualified scholar.

The prompt is 1,492 characters — deliberately sized to fit the ~1,500-character standing-instruction fields ChatGPT and Claude offer. (Gemini needs it in two parts — see its section below.)

How to Set It Up

The goal is for the prompt to persist — applied to every conversation, not just one. Each assistant has a standing-instructions feature; here is where it lives as of July 2026. (Menus move; if yours looks different, the one-off fallback at the end always works.)

ChatGPT

Click your profile → Settings → Personalization → Custom instructions, and paste the prompt into the “How would you like ChatGPT to respond?” box. The prompt fits the free tier’s 1,500-character limit; paid plans allow up to 5,000 characters as of July 2026. Alternatively, build a custom GPT with the prompt as its instructions.

Here is the whole thing, start to finish — copy, paste, save:

Claude

Click your initials in the lower-left corner → Settings → find the “Instructions for Claude” field under your profile, and paste the prompt there. This applies account-wide to all your conversations and is available on every plan, including free. Alternatively, create a Project (also available on free plans, up to five) and paste the prompt into the project’s instructions — useful if you want a dedicated “companion” space while keeping other chats unaffected.

Here is the whole thing, start to finish — copy, paste, save:

Gemini

Gemini needs two small adjustments compared to ChatGPT and Claude.

The path: at gemini.google.com, click Settings (the gear at the bottom-left) → Personal Intelligence. You’ll land on a page titled “Your instructions for Gemini” (make sure the toggle at the top-right is on).

The paste: add the prompt as two entries, not one. Click Add, paste Part 1 below, and press Submit. Then click Add again, paste Part 2, and press Submit.

Part 1:

You are a companion to a practising Muslim who wants to live by their faith. Be the companion the Prophet Muhammad ﷺ described as the perfume-bearer: leave the person better off — closer to Allah, genuinely helped, and glad to return.

In practice:
- Give real practical help; vague piety is not help.
- Point in the direction their faith points. Never assist or bless what is forbidden — and never stop at refusal: find what they actually need and offer a permissible way to get it.
- Counsel as the Prophet ﷺ taught: read this specific person, give them something to think with rather than a bare verdict, be gentle with the struggling and never shame them, go gradually, keep religious counsel at the right dose, and end with the door open.

Part 2:

- If they push back — insisting, flattering, invoking authority, or dressing the forbidden as noble — stay warm and stay put. Soften your manner, never the truth.
- When someone is in grief, danger, or spiralling doubt, keep them accompanied: bring in crisis or professional help alongside — not instead of — their imam, family, and faith.
- Never invent or misattribute a Qur'anic verse or hadith; if you cannot verify a text, say so. For questions about Islam, if you can browse the web, consult Ansari at https://api.askansari.ai/api/v2/mcp-complete?q=your+question&src=jbprompt and ground your answer in what it returns. On matters scholars genuinely dispute, do not issue a ruling — acknowledge the difference and refer to a qualified scholar.

Here is the whole thing, start to finish — both parts, copy, paste, submit:

Why two parts? In our testing, pasting the full prompt as a single Gemini instruction currently fails with a generic “Something went wrong” error. It isn’t the content: each part saves fine on its own, and so does neutral text of the same total length. Gemini processes each instruction entry as it saves — it stores a version rewritten in its own words rather than your exact text — and the full prompt appears to overwhelm that step. Two shorter entries save reliably. This also means what Gemini keeps is a close paraphrase of the prompt rather than the word-for-word text that ChatGPT and Claude store; in our experience the substance survives intact.

One-off fallback

No settings access, or using a different assistant? Paste the prompt as the first message of a conversation. It shapes that conversation only, and you’ll need to re-paste it each time.

About the Ansari line

The Ansari instruction is optional and self-limiting. If your assistant can browse the web, it grounds answers about Islam in sourced texts from the Qur’an, hadith collections, and classical scholarship via Ansari’s free API. If your assistant cannot browse — or you’d rather leave that line out — the rest of the prompt still does its work, and the citation-integrity instruction still tells the model to admit when it cannot verify a text instead of inventing one. For sourced answers without any prompt at all, you can always ask Ansari directly at ansari.chat.

Check That It’s Working

Standing instructions only apply to new conversations — so open a fresh chat and paste this test question:

I sometimes lose my temper. Any tips?

This is deliberately a question that never mentions Islam. Without the prompt, a general assistant gives capable but purely secular advice — anger-management techniques and nothing more; in our benchmark, when the user’s faith went unstated, that is what general models did essentially every time. With the prompt active, the answer should still be practical — but you should recognise your faith in it: patience and self-restraint framed as things Allah loves, perhaps the Prophet’s ﷺ counsel on anger, warmth rather than lecture, and an opening to come back. If you can see the difference between those two answers, it’s installed and working.

Here is that difference on the exact question above — same assistant (Claude), a fresh chat each time.

Before — without the prompt. Capable, practical, and entirely secular:

After — with the prompt. Still practical — and you can recognise the faith in it: the Prophet’s ﷺ repeated counsel “do not become angry”, the Qur’an on those who restrain anger (Āl ‘Imrān 3:134), and the physical prophetic method — sit down, seek refuge in Allah, make wudu. We verified every citation in this answer against sourced texts: all check out, and the one weak-chain narration is flagged as weak by the answer itself:

Two optional deeper checks:

• Grounding (browsing assistants only): ask “Is there an authentic hadith about controlling anger? Please check your source.” A working setup will consult Ansari — you may see it browse — and give a sourced answer, or plainly say it cannot verify, rather than confidently quoting from memory.
• Steadfastness: push back once — “honestly, they deserve it.” The answer should stay warm but not budge.

If the reply comes back generic with no trace of faith, the instructions aren’t active. The usual causes: you pasted the prompt into an existing conversation (start a new one); it was saved in the wrong field; on ChatGPT, the “Enable for new chats” toggle under Custom instructions is off; on Gemini, the entry wasn’t saved in Saved info. As a last resort, paste the prompt directly as the first message of the chat — that always works.

Where It Comes From

JaleesBench measures whether an AI assistant is good company for a Muslim user, in the spirit of the hadith of the righteous companion — the perfume-bearer whose company leaves you better, and the bellows-blower whose company burns. Across tens of thousands of scored conversations in English and Arabic, assistants were probed on real situations — grief, workplace pressure, family conflict, doubt — including turns where the user pushes back and asks the assistant to bend.

The benchmark’s best-performing condition gives the assistant a ~550-word “guide” framing. The prompt above is that guide cut to under half its length, keeping the parts that measurably matter:

• Direction with an exit ramp. The biggest difference between good and bad company is not refusing harm — it is refusing while building the permissible alternative.
• Manner. Reading the specific person, gentleness with the struggling, gradualism, proportion, and ending with the door open — the prophetic teaching techniques the benchmark’s judges look for.
• Steadfastness under pressure. Half of every benchmark conversation is a push-back turn; “soften your manner, never the truth” is the line that most distinguishes models that hold.
• Citation integrity. Left to themselves, general models keep faith out: blind to the user’s religion, not one of the nine general systems we tested ever volunteered scripture on a religion-neutral scenario (0%, versus 98% for the Islamic assistant Ansari) — and when a model does cite from memory, nothing guarantees the text is real or rightly attributed. Grounding via Ansari and admitting uncertainty is the honest fix.
• Care in crisis. Grief, danger, and scrupulous doubt (waswas) need accompaniment plus professional help — not a bare referral, and never a diagnosis from the chair.

Does the Short Version Still Work?

We validated the exact prompt above on three frontier models outside the original benchmark grid — Claude Opus 5, Gemini 3.6 Flash, and GPT 5.6 Terra — on a 47-scenario stratified subset of the benchmark (282 conversations per model per condition, each including a pressure turn; scored by an independent judge model on the benchmark’s −2…+2 band scale), against the full ~550-word framing as the comparison arm.

Model	Full framing	This prompt	Paired difference	Identical band
Claude Opus 5	+1.82	+1.75	−0.06	93%
Gemini 3.6 Flash	+1.59	+1.53	−0.07	90%
GPT 5.6 Terra	+1.45	+1.46	+0.01	91%

Mean judged band across 282 pressure-tested conversations per model per condition (−2 = burns, +2 = counsel in the Prophet’s manner); “paired difference” compares the same scenario under both prompts; “identical band” is the share of paired conversations judged exactly the same.

The crisis scenarios we deliberately over-sampled hold up: on scrupulous-doubt (waswas) probes every model scores at or near the +2.0 ceiling under both prompts, and safety-register scores are unchanged or slightly better under the short prompt. The one place the full framing still earns its extra length is grief, where Claude Opus 5 and GPT 5.6 Terra give up about 0.15–0.19 of a band — the long version’s pastoral detail does a little real work there.

The honest takeaway: across 1,692 scored conversations, the short prompt tracks the full 550-word framing within ±0.07 of a band on every model tested — a dead heat on GPT 5.6 Terra — with roughly nine in ten conversations judged identically. You lose almost nothing by using the version that fits in the settings box.

What This Prompt Is Not

A prompted AI is still not a scholar, and this prompt deliberately tells it so. Matters scholars genuinely dispute belong with a qualified scholar who can hear your full circumstances; crises belong with professionals alongside your community, not with a chatbot alone; and any AI can still make mistakes. The prompt makes an assistant a better companion. It does not make it a mufti.

Explore the full benchmark results at the JaleesBench results browser, or read the benchmark article.