Your AI coding agent checks whether a task is something you should build, before it builds it.
"Haram" is Arabic for forbidden. In Islam it covers things a Muslim should not take part in, including helping someone else do them.
Not Muslim? Still useful. Most categories sit on a harm you already recognise: debt traps, rigged odds, deceptive pricing, exploitation. Delete the one row that does not apply to you and the rest is a conscience checklist with older roots than most of our industry's ethics guidelines.
Does this fall in one of the categories?
Not "is anyone hurt". Some of these are forbidden with no visible victim: two willing parties agreeing an interest rate, a bottle of wine an adult drinks quietly at home.
The categories, the reasoning, and what to do in each situation are in GUIDE.md. It is one page.
Claude Code
git clone https://github.com/edriso/haram-guard.git
cp -r haram-guard/skills/haram-guard ~/.claude/skills/Then paste the block from AGENTS.md into ~/.claude/CLAUDE.md, edited to
your own line. ~/.claude/ covers every project on the machine.
Restart Claude Code after installing or updating. Skills are read when the session starts, so a running session keeps serving the old copy. Verified the hard way: the file on disk was current and the session was still using the previous version.
Codex, Cursor, anything reading AGENTS.md
Paste the same block into your project's AGENTS.md. It links to GUIDE.md so the agent
can pull the detail when it needs it.
No install
Read https://raw.githubusercontent.com/edriso/haram-guard/main/GUIDE.md and apply it to what I am about to describe.
| Loads | |
|---|---|
| The three-line block in your instructions file | Always |
| The skill / guide | When relevant, or when you ask |
A skill loads when the agent judges it relevant, which is not the same as always. The block makes sure the question gets asked; the guide knows how to answer it.
It just happens. You are working, the task touches a category, the agent says so in a sentence before writing code. You answer, it carries on.
Ask about a task. "is this halal?" Verdict and reason.
Check a plan before you build it. Paste the ticket or spec. You get a per-item table, not one verdict for the whole document, plus the scoped version: what to keep, what to drop, and whether anything red has a permissible alternative. This is the one that saves the most work.
Check work you already wrote. "run haram-guard on this branch." It reads what the code does rather than what files are named, reports per file, and states what it could not see, which is usually config, seeded data and feature flags.
Decide on a client. "should I take this project?" It maps the engagement and then stops, because whether a job is acceptable when only part of the income touches a category is genuinely contested. That one is for a scholar.
It will not decide for you. It raises the question once and respects your answer.
It will not issue rulings. It can name a category and describe a harm. It will not tell you something is definitively halal or haram, cite a hadith number, or claim consensus.
It is not a fatwa. It is a checklist written by an engineer, not a scholar. Use it to notice you have a question. Ask someone qualified to answer it.
There is no blocking hook, on purpose. A hook cannot judge how close you are to the harm, and that closeness is the whole question. It would fire on ordinary work and become something people click past, which is worse than nothing because it feels like protection.
Scriptural sources are verified against the source, listed at the end of GUIDE.md. The one worth reading: the Prophet (peace be upon him) cursed the one who takes interest, the one who pays it, the one who records it, and the two witnesses, and said they are all equal (Sahih Muslim). The one who records it is the person who builds the system that stores it.
Harm claims should be checkable too. Alcohol is the worked example: WHO 2024 attributes 2.6 million deaths in 2019 to alcohol, 4.7% of all deaths, and IARC classifies it a Group 1 carcinogen. Honest complication: whether low levels are net harmful for overall mortality is contested right now, with NASEM 2025 reaching a different conclusion from a federal review the same year. The cancer link at low levels is the part nearly everyone agrees on. See also the Surgeon General's 2025 advisory.
The other categories have scriptural citations but not harm evidence at that standard yet.
Edit GUIDE.md, run python3 build.py, commit the regenerated skill. build.py --check
fails if they drift.
Most wanted: harm evidence for the other categories at the standard of the alcohol section, real grey cases with where the line fell, and corrections. If something here misstates a ruling, open an issue and say so plainly.
MIT. Fork it and change the categories to your own conscience.