Home › Rules › Inbox / email AI agent

Inbox / email AI agent rules: AGENTS.md, CLAUDE.md & Cursor rules

Free to copy. File formats checked against the official docs on Oct 19, 2026 (sources).

• AI agents & apps
• Claude Code
• OpenAI Codex
• Cursor
• GitHub Copilot

An email agent that can delete or send is one bad prompt away from a very bad day. These rules keep yours to labeling and drafting, tested on 20 emails first.

Stack these rules assume: Python or Node, Gmail API (OAuth), an LLM API for classification. Different stack? Customize it in the generator.

Quick start

1. Copy the AGENTS.md below into your project's root folder.
2. Fill in the three [bracketed] lines under "Project". Delete any rule that doesn't fit.
3. Start your first chat with: "Read AGENTS.md and follow it for this whole session."

AGENTS.md (works across tools)

Read by OpenAI Codex, Cursor, GitHub Copilot, Claude Code (when there's no CLAUDE.md) and many other agents.

# AGENTS.md: Inbox / email AI agent

Instructions for AI coding agents working on this project. Replace anything in [brackets].

## Project
- What it is: [one sentence: what you are building and who it is for]
- Stack: Python or Node, Gmail API (OAuth), an LLM API for classification
- Version 1 only needs to: [the ONE thing it must do]

## Commands
- Test on a small batch: `python triage.py --dry-run --limit 20`
- Run for real (labels only): `python triage.py --limit 50`

## Project rules
- The agent may add labels and create drafts. It must never delete, archive, send, or forward email unless I add that later on purpose.
- Ask for the narrowest Gmail OAuth scope that does the job.
- credentials.json and token.json (or equivalents) are in .gitignore and never printed.
- Every run supports --dry-run and --limit. Default to a small limit.
- The classifier returns strict JSON with a fixed label list. Anything unparseable goes to a "Needs review" label.
- Set a max number of LLM calls per run and print the count at the end.

## How to work
- Do one task per request. If my request is really several tasks, list them and do only the first.
- Before writing code, give a 3-line plan: what you'll change, which files, and how I'll test it.
- Make the smallest change that works. Don't refactor, rename, or reformat code I didn't ask about.
- Ask before adding a dependency, deleting or moving files, changing the data/database structure, or touching more than 3 files.
- If you're not sure a library, API, or setting exists in the version we use, say so and check the installed package or the docs instead of guessing.
- Never say something works unless you ran it or I confirmed it. Report errors exactly.
- Secrets (API keys, tokens, passwords) live only in environment variables or a .env file listed in .gitignore. Never print, hardcode, or commit them.
- Never run destructive commands (rm -rf, git reset --hard, force-push, dropping tables) without asking first.

## Finish every task with
1. What changed (list the files).
2. How to run or test it in under a minute, and what I should see.
3. A suggested git commit message.

Why each rule exists

• Inbox / email AI agent ruleThe agent may add labels and create drafts. It must never delete, archive, send, or forward email unless I add that later on purpose.Why: Labels are reversible. Deleted or sent email isn't.
• Inbox / email AI agent ruleAsk for the narrowest Gmail OAuth scope that does the job.Why: Broad scopes are more dangerous if a token leaks, and harder to get approved.
• Inbox / email AI agent rulecredentials.json and token.json (or equivalents) are in .gitignore and never printed.Why: Those files are the keys to your inbox.
• Inbox / email AI agent ruleEvery run supports --dry-run and --limit. Default to a small limit.Why: You want to see what it would do to 20 emails before it touches 2,000.
• Inbox / email AI agent ruleThe classifier returns strict JSON with a fixed label list. Anything unparseable goes to a "Needs review" label.Why: LLMs sometimes return junk. A safe fallback beats a wrong label.
• Inbox / email AI agent ruleSet a max number of LLM calls per run and print the count at the end.Why: Runaway loops on an LLM API cost real money.
• How to workDo one task per request. If my request is really several tasks, list them and do only the first.Why: Big multi-part requests are where AI tools change 12 files and you can't tell which change broke things.
• How to workBefore writing code, give a 3-line plan: what you'll change, which files, and how I'll test it.Why: A plan you can read in 10 seconds catches the wrong approach before it costs you an hour or a pile of credits.
• How to workMake the smallest change that works. Don't refactor, rename, or reformat code I didn't ask about.Why: Drive-by "cleanups" are the #1 cause of "it was working yesterday".
• How to workAsk before adding a dependency, deleting or moving files, changing the data/database structure, or touching more than 3 files.Why: These are the changes that are hard to undo. One quick question is cheaper than a rollback.
• How to workIf you're not sure a library, API, or setting exists in the version we use, say so and check the installed package or the docs instead of guessing.Why: AI tools confidently invent functions and config options. Making it check is the cure.
• How to workNever say something works unless you ran it or I confirmed it. Report errors exactly.Why: "Should work now" isn't a test result. You want facts, not optimism.
• How to workSecrets (API keys, tokens, passwords) live only in environment variables or a .env file listed in .gitignore. Never print, hardcode, or commit them.Why: Leaked keys get found by bots within minutes of a public push. This rule is the cheapest insurance you'll ever buy.
• How to workNever run destructive commands (rm -rf, git reset --hard, force-push, dropping tables) without asking first.Why: One bad command can delete the only copy of your work.
• Finish every task withWhat changed, how to test it, and a commit message.Why: This turns every task into a checkpoint: you test it, you commit it, and you always have a last good version to go back to.

Use it as CLAUDE.md (Claude Code)

Pick one:

1. Do nothing. Current Claude Code versions read AGENTS.md when the project has no CLAUDE.md. Run /memory in a session to check it loaded.
2. Import it (best if you want Claude-only extras, or you're on an older Claude Code). Create CLAUDE.md next to AGENTS.md:

@AGENTS.md

## Claude Code
- Use plan mode before any change that touches more than 3 files.

3. Claude only? Save the same AGENTS.md text as CLAUDE.md instead. Don't keep two separate copies; they drift apart.

Heads-up: if a project has both files and CLAUDE.md doesn't import AGENTS.md, Claude Code reads only CLAUDE.md by default. More in AGENTS.md vs CLAUDE.md.

Use it as Cursor rules (.cursor/rules/*.mdc)

Cursor reads AGENTS.md too, so the file above already works. If you prefer Cursor's own Project Rules, split it into two .mdc files. The .mdc extension and frontmatter are required; Cursor ignores plain .md files in .cursor/rules.

---
description: Project overview and working rules for Inbox / email AI agent
alwaysApply: true
---

# Inbox / email AI agent

Instructions for AI coding agents working on this project. Replace anything in [brackets].

## Project
- What it is: [one sentence: what you are building and who it is for]
- Stack: Python or Node, Gmail API (OAuth), an LLM API for classification
- Version 1 only needs to: [the ONE thing it must do]

## Commands
- Test on a small batch: `python triage.py --dry-run --limit 20`
- Run for real (labels only): `python triage.py --limit 50`

## How to work
- Do one task per request. If my request is really several tasks, list them and do only the first.
- Before writing code, give a 3-line plan: what you'll change, which files, and how I'll test it.
- Make the smallest change that works. Don't refactor, rename, or reformat code I didn't ask about.
- Ask before adding a dependency, deleting or moving files, changing the data/database structure, or touching more than 3 files.
- If you're not sure a library, API, or setting exists in the version we use, say so and check the installed package or the docs instead of guessing.
- Never say something works unless you ran it or I confirmed it. Report errors exactly.
- Secrets (API keys, tokens, passwords) live only in environment variables or a .env file listed in .gitignore. Never print, hardcode, or commit them.
- Never run destructive commands (rm -rf, git reset --hard, force-push, dropping tables) without asking first.

## Finish every task with
1. What changed (list the files).
2. How to run or test it in under a minute, and what I should see.
3. A suggested git commit message.

The project rules only attach when a matching file is in the chat (globs with alwaysApply: false). Want them on all the time? Change it to alwaysApply: true.

---
globs: **/*.py, **/*.js, **/*.ts
alwaysApply: false
---

# Project rules: Inbox / email AI agent

- The agent may add labels and create drafts. It must never delete, archive, send, or forward email unless I add that later on purpose.
- Ask for the narrowest Gmail OAuth scope that does the job.
- credentials.json and token.json (or equivalents) are in .gitignore and never printed.
- Every run supports --dry-run and --limit. Default to a small limit.
- The classifier returns strict JSON with a fixed label list. Anything unparseable goes to a "Needs review" label.
- Set a max number of LLM calls per run and print the count at the end.

Use either AGENTS.md or these .mdc files, not both, so you don't send the same rules twice. See Cursor rules vs AGENTS.md.

Related

• Best AI tool to automate email triage (which tool to use, real prices)
• AGENTS.md, CLAUDE.md and Cursor rules formats explained