Reminders Bridge

Talk to Claude on your phone. It builds on your Mac.

Claude's voice mode is great for thinking through ideas — real-time, hands-free, back-and-forth conversation. Claude Code on your Mac is great for building things. But they don't talk to each other.

Reminders Bridge connects them using Apple Reminders as a message bus. You talk through an idea with Claude on your phone, Claude drops a prompt into your Reminders, and Claude Code on your Mac picks it up, matches it to a project, and dispatches an agent to build it. Results go back to Reminders so you can check from your phone.

I used this to build Infinite Orbit — a gravity simulator game — mostly by voice while walking around outside.

How It Works

┌─────────────────────┐          ┌──────────────────┐          ┌─────────────────────┐
│   Claude Voice       │          │  Apple Reminders   │          │    Claude Code        │
│   (iPhone)           │          │  (iCloud sync)     │          │    (Mac)              │
│                      │          │                    │          │                       │
│  You brainstorm an   │  ──1──▶  │  "Claude Inbox"    │  ──2──▶  │  /reminders skill     │
│  idea with Claude    │          │  gets a new        │          │  detects new item     │
│  using voice mode    │          │  reminder          │          │                       │
│                      │          │                    │          │  Matches to a project │
│                      │          │                    │  ◀──3──  │  folder, dispatches   │
│  Check results on    │  ◀──4──  │  "Claude Output"   │          │  a background agent   │
│  your phone          │          │  gets the result   │          │                       │
└─────────────────────┘          └──────────────────┘          └─────────────────────┘

1. You + Claude Voice — Talk through your idea on your phone. Let Claude ask questions, suggest features, play devil's advocate.
2. Claude writes a prompt — When the idea is ready, Claude creates a reminder in "Claude Inbox" with the full prompt.
3. Claude Code picks it up — The /reminders skill (running on a loop) detects the new item, matches it to one of your project folders, and dispatches a background agent.
4. The agent builds — It reads your project context (CLAUDE.md, SESSION_LOG.md), does the work, and writes a summary back to "Claude Output."
5. You check from your phone — Ask voice Claude to read your Claude Output reminders, or just open the Reminders app.

Requirements

• macOS (uses AppleScript + Apple Reminders)
• Claude Code CLI installed and authenticated
• Claude Pro or Max plan (for voice mode + CLI agents)
• iPhone with Claude app (for voice conversations)
• iCloud Reminders sync enabled on both devices

Setup

1. Install

git clone https://github.com/brianharms/reminder-watch.git
cd reminder-watch
./install.sh

2. Create the Reminder Lists

Open Apple Reminders and create two lists (names must match exactly):

• Claude Inbox — where prompts from your phone land
• Claude Output — where results go back

3. Start Watching

Open Claude Code and run:

/loop 5m /reminders

The first time, it'll ask where your projects folder is. Just tell it in plain English — "it's on my desktop in the projects folder" works fine. After that, it watches silently.

Usage

Sending Tasks from Your Phone

Start a voice conversation with Claude and brainstorm your idea. When it's ready:

"Send that to Claude Code as a reminder in my Claude Inbox list. Use the title for a short description and the notes for a detailed prompt."

Tips for good prompts:

• Be specific about the project name. "Work on my gravity simulator" is better than "work on that game thing."
• Push Claude for detail. Voice Claude tends to summarize. Say: "Include everything we discussed — the physics, the controls, the visual style. Don't summarize."
• One task per reminder. Don't pack multiple unrelated tasks into one reminder.

Checking Results

From your phone:

• Ask voice Claude: "Check my Claude Output reminders and read them to me"
• Or just open the Reminders app

From Claude Code:

• Run /reminders to check manually

Voice Session Setup

Voice mode doesn't persist instructions across conversations. At the start of each voice session, say something like:

"When I tell you to send something to Claude Code, create a reminder in my Claude Inbox list. Title should be a short description, notes should be the full detailed prompt with all context."

How Project Matching Works

When a new reminder comes in, Claude Code lists your project folders and picks the best match based on the task title and notes:

• "Work on mac caps" → matches mac-caps/
• "Let's build a new weather app" → creates weather-app/ (new folder)
• "Add a feature to the portfolio site" → matches portfolio-site/

Be specific in your prompts. The more clearly you name the project, the better the match. If your folder is called recipe-app, say "work on recipe-app" not "work on that cooking thing."

If no existing folder matches, Claude creates a new one with a kebab-case name in your projects directory.

Configuration

After first run, your config lives at:

~/.claude/reminders-config.sh

It contains one line:

PROJECTS_DIR="/path/to/your/projects"

To change your projects folder, edit this file or delete it and run /reminders again.

What's Included

Installed File Locations

Troubleshooting

Reminders aren't being detected:

• Check that the list is named exactly Claude Inbox (capitalization matters)
• Make sure iCloud sync is working (changes on phone should appear on Mac within a few seconds)
• Try running the check script manually: ~/.claude/reminder-check.sh

Nothing happens when reminders come in:

• Make sure you have /loop running: /loop 5m /reminders
• Check that Claude Code CLI is authenticated: claude --version
• Try /reminders manually to see if it picks up the item

Wrong project matched:

• Be more specific in your reminder title — use the exact folder name if possible
• Check what folders exist in your projects directory

Keeping Reminders clean:

• Tell voice Claude to mark output reminders as complete after reading them
• Apple Reminders slows down with many incomplete items that have long notes

Uninstall

Delete these files:

rm ~/.claude/reminder-check.sh
rm ~/.claude/reminder-output.sh
rm -rf ~/.claude/skills/reminders/
rm ~/.claude/reminders-config.sh    # your config
rm ~/.claude/reminder-seen.txt      # tracking file

Background

I built this because I think best when I'm moving — walking, not sitting at a desk. The technology exists to have a voice conversation that develops an idea and hands it off to a coding agent. Nobody's shipped that as a native experience yet, so I hacked it together with Apple Reminders and a few shell scripts.

It's not elegant, but it works. I hope it inspires the right people to build this properly.

— Brian Harms

For AI coding agents

This is a small, dependency-free Bash + AppleScript project. macOS-only (it talks to Apple Reminders via osascript). The folder is named reminder-watch; the user-facing product name is Reminders Bridge.

Repo layout

• scripts/reminder-check.sh — Reads the Claude Inbox list for incomplete reminders, filters out IDs already recorded in ~/.claude/reminder-seen.txt, marks each new one complete (so it isn't reprocessed), and also reads the Claude Output list for flagged items. Emits pipe-delimited lines (NEW|title|notes, FLAGGED|title|notes); exits 1 when nothing is new.
• scripts/reminder-output.sh — Writes a reminder into the Claude Output list. Usage: reminder-output.sh "Title" "Body" [true|false] (third arg true sets priority to 1 / urgent). Called by dispatched agents to report results back to the phone.
• skill/SKILL.md — The /reminders Claude Code skill (frontmatter name: reminders). Orchestrates the whole loop: config check → run check script → match each new task to a project folder → dispatch a background agent → report. Read this first to understand the end-to-end flow.
• install.sh — Copies the two scripts to ~/.claude/ and the skill to ~/.claude/skills/reminders/, touches the seen file, and prints next steps. No daemon, no build step.
• README.md — User-facing docs (setup, usage, project-matching, troubleshooting).
• banner.png, infinite-orbit-screenshot.png — README images.
• docs/ — empty placeholder directory.
• .gitignore, LICENSE (MIT), .gitignore.

Entry points

Start with skill/SKILL.md (the control flow), then scripts/reminder-check.sh (the read side) and scripts/reminder-output.sh (the write side). There is no application binary — the "runtime" is Claude Code executing the skill on a loop.

Build / run / test locally

There is nothing to compile. To exercise it:

1. In Apple Reminders, create two lists named exactly Claude Inbox and Claude Output.
2. ./install.sh (copies into ~/.claude/), or run the scripts straight from scripts/ for testing.
3. Add a reminder to Claude Inbox, then run bash scripts/reminder-check.sh — expect a NEW|... line and exit 0; run it again and expect exit 1 (already seen / marked complete).
4. Test the write side: bash scripts/reminder-output.sh "Test" "Hello" true — a flagged reminder should appear in Claude Output.
5. End-to-end inside Claude Code: /reminders once (configures the projects folder), then /loop 5m /reminders.
   First run with a fresh seen file will mark existing inbox reminders complete — test with disposable reminders.

Invariants — do not break

• macOS-only. All inbox/output access goes through osascript. Don't replace AppleScript with anything that isn't present by default on macOS, and keep the uname == Darwin preflight in install.sh.
• Exact list names. The literal strings "Claude Inbox" and "Claude Output" are the protocol contract between phone-Claude and desktop-Claude. Don't rename them or make them configurable without updating the README, skill, and both scripts together.
• Pipe-delimited protocol tokens. reminder-check.sh emits NEW|... and FLAGGED|..., and SKILL.md parses exactly those prefixes. Keep the | field delimiter and \t-delimited AppleScript output (reminder id \t name \t body) in sync between the script and any consumer.
• Idempotency / seen-tracking. New reminders are recorded in ~/.claude/reminder-seen.txt and marked completed to prevent reprocessing. Preserve this dedupe behavior; reminder id (not name) is the identity key.
• Exit-code contract. reminder-check.sh returns 1 for "nothing new" and 0 when items exist; the skill relies on 1 to return silently. Don't change these semantics.
• AppleScript string escaping. reminder-output.sh escapes backslashes and double-quotes via escape_as() before interpolating into the AppleScript literal. Any new user-supplied string injected into an osascript heredoc must be escaped the same way.
• Agent guardrails in the dispatch prompt. The SKILL.md agent prompt forbids deleting files, git push/deploy, posting to external APIs, and touching projects outside the resolved PROJECT DIR. Keep these guardrails when editing the dispatch template.
• Secrets stay out of git. .gitignore excludes .env*, certs/, *.pem/*.key/*.p12, credentials.json, *-seen.txt, *.log, plus internal dev notes (CLAUDE.md, SESSION_LOG.md, .claude/). Never commit these or the user's seen-tracker.

Placeholdered / personal config — keep generic

• No hardcoded projects path. The projects directory is never baked in. It is resolved at first run and written to ~/.claude/reminders-config.sh as a single line PROJECTS_DIR="...". Keep this parameterized — don't reintroduce a personal absolute path (e.g. anyone's ~/Desktop/...) into the scripts or skill.
• All state lives under ~/.claude/ via ${HOME} (reminder-seen.txt, reminders-config.sh, installed scripts/skill). Keep paths $HOME-relative, not user-specific.
• The only personal references that should remain are the author attribution and example URLs in README.md/script header comments (github.com/brianharms/reminder-watch, the Infinite Orbit links, MIT copyright). Example project names in the README (mac-caps, weather-app, portfolio-site) are illustrative only — don't treat them as real paths.