⌃⌥Spacefrom any app
One notch. Every prompt.
Access prompt optimization directly from your MacBook's notch. On a display without one, Refyn waits at the top center.
The prompt optimizer in your notch
Optimized prompts.Completely free.
Meet Refyn. The free AI prompt optimizer that lives in your Mac's notch. Turn rough ideas into precise, actionable prompts in seconds.
Free to use · macOS 14 or later · Apple Silicon
Demo
Try the notch.
This is Refyn's real interface with a prewritten example. Nothing is sent anywhere. In the app, your Mac generates the prompt in a few seconds.
Click the notch to open Refyn.
Examples
Same request. Far better brief.
Before Refyn
checkout keeps failing sometimes when ppl use a coupon?? not sure why. can u fix it
What Refyn added
- A way to reproduce it
- A hard stop on guessing
- A regression test
After Refyn
## Objective
Find and fix the intermittent checkout failure that occurs when a coupon is applied.
## Investigation
- 1.Reproduce with a valid, an expired, and an already-used coupon. Record which combinations fail.
- 2.Trace the coupon from input to order total. Look for race conditions, stale cart state, and unhandled API errors.
- 3.State the root cause before editing any code.
## Constraints
- –Keep the fix inside the coupon and checkout flow.
- –Don't change pricing rules or which coupons are valid.
## Verification
- –Add a regression test that fails before the fix and passes after it.
- –Confirm checkout still completes without a coupon.
Before Refyn
make the settings page look nicer, its kinda ugly and cramped
What Refyn added
- What “nicer” means
- Limits on the redesign
- Checks at every width
After Refyn
## Objective
Improve the visual hierarchy and spacing of the settings page without changing what it does.
## Tasks
- 1.Group related settings under clear section headings.
- 2.Apply the existing spacing and type tokens so sections have room and a consistent rhythm.
- 3.Align labels, help text, and controls so the page scans in one pass.
## Constraints
- –Reuse existing components and design tokens.
- –Add no new dependencies.
- –Every setting keeps its current behavior.
## Verification
- –Check the page at mobile, tablet, and desktop widths.
- –Tab through every control and confirm the focus order and focus ring.
Before Refyn
add dark mode
What Refyn added
- The actual scope
- An order of work
- Edge cases to test
After Refyn
## Objective
Add a dark theme people can choose, following the system setting by default.
## Tasks
- 1.Find how colors are defined today and move hard-coded values into semantic tokens.
- 2.Define dark values for every token.
- 3.Add a System, Light, and Dark setting and persist the choice.
- 4.Apply the theme before first paint so the page never flashes.
## Constraints
Change color only. Leave layout and copy alone.
## Verification
- –Confirm text meets WCAG AA contrast in both themes.
- –Reload and confirm the choice persists.
- –Change the system appearance and confirm System mode follows it.
Before Refyn
this auth file is like 900 lines and a mess, clean it up but dont break stuff
What Refyn added
- A definition of “don't break”
- Safety tests first
- Small, reversible steps
After Refyn
## Objective
Split the authentication file into focused modules with no change in behavior.
## Steps
- 1.Map its responsibilities, public exports, and callers.
- 2.Add tests around current behavior wherever coverage is missing.
- 3.Extract one responsibility at a time and run the tests after each move.
- 4.Keep the original path re-exporting the public API.
## Constraints
- –The public API and session format stay the same.
- –No new features or dependency upgrades.
## Verification
- –Run the full test suite.
- –Exercise login, logout, and token refresh by hand.
Before Refyn
dashboard takes forever to load, make it fast
What Refyn added
- Measure before changing
- Fix what's actually slow
- Proof it got faster
After Refyn
## Objective
Reduce the dashboard's load time by fixing its largest measured bottlenecks.
## Investigation
- 1.Record a baseline on a production-sized account: time to first data, largest contentful paint, and the request waterfall.
- 2.Rank the causes: slow queries, over-fetching, bundle size, and unnecessary re-renders.
- 3.Fix the top two, one at a time, and measure after each.
## Constraints
- –Displayed data and permissions stay the same.
- –Don't add a cache without defining when it invalidates.
## Verification
- –Report before and after numbers for the same account.
- –Confirm the dashboard still shows fresh data.
Features
Small enough to forget. Ready when you need it.
Claude Code
Codex
Cursor
Built for coding agents.
Turn vague requests into precise instructions for Codex, Claude Code, and other coding assistants.
Running locallyQwen3 · on-device
Local AI optimization.
Qwen3 runs on your Mac through MLX. No API keys, no cloud inference, no per-prompt charges, and it works offline after one model download.
Fix coupon checkout failure
Split auth into modules
Add dark theme tokens
Your prompts, organized.
Save the prompts worth keeping and search your history. Add a workspace and notes, and Refyn brings in only the relevant context.
Listening…Hold ⌃⌥D
Voice to prompt.
Hold a shortcut and say what you need. Refyn transcribes it on your Mac and leaves it in the editor. Download a speech model once from Settings.
Needs your input2 questions
Which screen feels slow?
Your answer (optional)
Asks before it guesses.
When a request is missing something important, Refyn asks up to three quick questions instead of inventing the answer.
How it works
Three steps to a better prompt.
- my app is slow and buggy, fix it
Write anything.
Enter a rough prompt directly into Refyn.
- Running locallyQwen3 · on-device## ObjectiveFind what is making the app slow## Investigation1. Reproduce each problem2. Profile before changing code## Verification
Refyn it.
Let the local AI improve clarity, structure, and technical specificity.
- For Claude CodeCopiedAgent ready.> ## ObjectiveFind what is making the app slow…[pasted 13 more lines]
Ship faster.
Copy the optimized prompt into your favorite coding agent.
What happens when you press Optimize.
Every step runs on your Mac. There are no API keys and no cloud inference. Refyn only goes online to download the models.
On your MacWorks offline after setup
Read the request
Refyn works out the intent and task type. If something important is missing, it asks up to three quick questions.
Gather context
From the workspace you choose, it pulls only relevant component summaries, symbols, paths, and notes. Never whole files, and never secrets.
Generate with Qwen3
Small requests are handled instantly by local rules. The rest go to Qwen3, running on your Mac through MLX, which drafts a structured plan.
Check the work
Code references it can't confirm are dropped, and unsupported names, paths, and numbers are caught. If a plan fails, built-in rules take over and Refyn tells you.
Write the prompt
Refyn composes the final prompt on your Mac, ready to copy into your agent or save for later.
- Main model
- Qwen3 4B · 4-bit
- Runs on
- MLX · Apple Silicon
- One-time download
- 2.3 GB
- Optional fast model
- Qwen3 1.7B · about 1 GB
- Cost per prompt
- Nothing
Your next prompt starts here.
Download for macOS- Free to use
- macOS 14 or later
- Apple Silicon
- About 8 GB of free memory recommended
- 2.3 GB model download on first setup