Lumi Documentation

Installation

1 Download

Download Lumi from our store. You'll receive a .zip file.

2 Open the installer

macOS blocks unsigned downloaded apps on double-click. Use one of these methods:

Terminal (one command):

xattr -cr ~/Downloads/Lumi-v1.0.0/ && open ~/Downloads/Lumi-v1.0.0/Install\ Lumi.command

Or right-click method (no Terminal):

1. Right-click (or Ctrl+click) on Install Lumi.command
2. Select Open from the context menu
3. Click Open in the security dialog

3 Done

The installer copies Lumi.app to /Applications and launches it. Subsequent launches work normally — macOS remembers your choice.

Or skip the installer: drag Lumi.app to Applications, then right-click → Open.

4 Grant permissions

Lumi needs Screen Recording permission to capture system audio. macOS will prompt you. After granting, restart the app (required by macOS on first grant).

First Launch

1. The onboarding wizard guides you through setup
2. Enter your STT and AI provider keys
3. Grant Screen Recording permission
4. Press Cmd+Shift+O to show the overlay — you're ready!

Provider Setup

Lumi can use OpenAI Whisper or Deepgram for speech-to-text, and OpenAI, Claude, Groq, Gemini, or Ollama for text responses. Screenshot analysis currently uses OpenAI vision.

1. Open Lumi Settings and choose your STT backend and AI backend
2. Add the required provider keys, such as OpenAI or Deepgram
3. If you want screenshot analysis, add an OpenAI key for vision

Start Listening

Click Start Listening or press Cmd+Shift+Space to begin capturing audio from your video call, YouTube, or any app playing sound.

1. Transcript appears in real-time in the left panel
2. Click Ask AI anytime to get an answer
3. Click End Session when done — see your performance metrics

Action Buttons & Ask AI

The top bar has one-click action buttons that execute immediately:

You can also:

• Type a question — "summarize what was discussed", "explain that concept"
• Use slash commands — /recap, /code, /followup, /answer
• Select from More menu — all stored modes and custom prompts

If screenshots are queued, actions automatically route through GPT-4o vision.

Response Transforms

After getting an AI response, use the transform buttons below it:

• Shorten — condense to essential points
• Expand — add more detail and examples
• Bullets — reformat into bullet points
• Copy — copy response to clipboard

Transforms are origin-aware: if the answer came from screenshots, the transform stays in the vision path. They don't pollute conversation history.

Smart Intent Routing

Lumi automatically classifies your query and adapts the AI system prompt. It detects 7 intents: answer, clarify, code, brainstorm, recap, follow-up, system design. This happens locally — no extra API call.

Knowledge Base (Playbooks)

Upload your documents so the AI answers with your specific details — not generic advice.

Setting up a playbook:

1. Settings → Playbooks tab → New Playbook
2. Click Add File — select multiple files at once (PDF, TXT, MD, RTF). Image-based/scanned PDFs are supported via OCR.
3. Click Set Active

What to upload:

Lumi intelligently finds the most relevant parts of your documents for each question — even if you upload 50+ pages of notes.

Custom Modes

Each Mode sets the AI system prompt and context strategy. The active Mode controls how Answer, Recap, and Follow Up behave.

Built-in modes:

• Interview — auto-detects question type (behavioral → STAR, coding → solution, system design → structured breakdown)
• Meeting — summaries with action items and suggested responses
• Coding — code review, bug detection, complexity analysis
• General — adaptive format based on content

Create your own:

1. Settings → Modes tab → New Mode
2. Name it, pick an icon, write the system prompt
3. Choose a context strategy (Interview-focused, Meeting-full, Coding-screen, Balanced)

Edit built-ins:

Click the pencil icon next to any built-in mode to customize its prompt. Click the reset icon to restore factory defaults at any time.

Quick Actions

One-click action buttons at the top of the overlay:

• Answer — generate an answer using transcript + playbook + conversation history
• Recap — summarize what's been discussed so far
• Follow Up — suggest questions to ask
• Shorten — condense the last AI response to key points
• Expand — add more detail and examples to the last response

If screenshots are queued, Answer/Recap/Follow Up automatically route through GPT-4o vision.

AI Playbook Synthesis

Click Re-organize on a playbook to have AI merge all documents into one structured markdown (Profile, Experience, Key Achievements, Skills, Projects, Notes). The synthesis becomes the primary context source — AI answers draw from this organized summary first.

Contextual Hints

Add technical terms and jargon so the speech engine recognizes them correctly. Hints improve transcription accuracy across all STT providers.

Organize hints by category (tech, project, company). Loaded once per session — no per-chunk overhead.

1. Settings → Hints tab
2. Add categories and terms (e.g., "Kubernetes", "gRPC", "ERC-3643")
3. Click Save — hints are used the next time you start listening

Screenshot Queue

Capture up to 5 screenshots that are sent to AI together in one call.

1. Press Cmd+Shift+S (or click Capture) — screenshot is added to the queue
2. Thumbnails appear in the capture panel with a count (e.g., "Screenshots 3/5")
3. Capture more screenshots to build context (e.g., multiple slides, code files, whiteboard)
4. Press Cmd+Shift+Return (or click Ask AI) — all images are sent to GPT-4o vision

Managing screenshots:

• Remove one — click the X on any thumbnail
• Undo/redo — arrow buttons in the screenshot bar
• Clear all — click the X button on the right
• Max 5 — oldest screenshot is dropped when you exceed the limit

Works for coding questions, whiteboard problems, slides, or any visual content.

Session Management

Lumi automatically saves your work so you never lose data — even if the app crashes.

How it works:

• Auto-save — sessions flush to disk every 30 seconds, or when memory exceeds 500KB
• Session picker on launch — see your recent sessions and resume any one, or start fresh
• Resume — pick up where you left off: transcript, AI conversation history, and playbook context are fully restored

Managing sessions:

Go to Settings → Sessions to:

• Resume — click to restore a previous session
• Rename — click the pencil icon to give sessions meaningful names
• Export — save as JSON for backup or sharing
• Import — load JSON files (supports multi-select)
• Delete — remove sessions you no longer need

Session Export

Press Cmd+Shift+E to save the entire session as a Markdown file:

• Full transcript with speaker labels
• All AI responses
• Screenshot analyses
• Session metadata (duration, mode, playbook)

Performance Metrics

Click End Session to see how you did:

• Talk ratio — your speaking time vs others
• Filler words — "um", "uh", "you know" count
• Response pace — how quickly you answered
• Questions answered — how many you responded to

Track your improvement across multiple sessions.

Stealth Mode

By default, Lumi is completely invisible to screen sharing — Teams, Meet, Zoom cannot see it.

The eye icon in the control bar toggles visibility:

• Eye-slash (grey) — invisible to screen sharing (default)
• Eye (red) — visible, for recording demos

Stealth toggle applies to all windows — the overlay, region selector, and performance view are all hidden from screen share simultaneously.

Updates & Licensing

Lumi checks for updates automatically on launch. When a new version is available, a banner appears at the top of the overlay.

• License key — enter your Gumroad license key in Settings → AI to unlock
• Signed updates — update files are cryptographically signed (Ed25519) to prevent tampering
• Auto update check — compares your version against the latest release

How We Protect Your Data

Your documents and credentials are protected with industry-standard encryption. Files stored on disk are unreadable without your machine.

• API keys — encrypted at rest, tied to your machine's hardware
• Playbook documents — encrypted at rest, stored as binary files
• No Lumi cloud storage — local files stay in ~/Library/Application Support/Lumi/
• Delete anytime — remove the folder and all data is gone

Privacy

• No analytics, no telemetry, no tracking
• No account required — no email, no signup
• Your documents never leave your machine
• AI queries are sent to the provider you choose (OpenAI, Anthropic, Groq, Gemini, or local Ollama)
• Audio is processed by your chosen STT provider (Whisper API or Deepgram), then discarded
• Ollama users: everything stays on your machine — no data sent anywhere

Testing & Verification

Lumi is backed by 455 automated tests across 40+ test suites, covering 70% of source files.

What we test:

• AI uses your documents — we upload a test resume and verify the AI response contains exact details like company names, project metrics, and tech choices
• Conversation context — a 7-round simulated interview verifies the AI references earlier answers when asked follow-up questions
• Transcription quality — real audio files are transcribed and checked for keyword accuracy
• Anti-hallucination (7 layers) — silence produces zero output; garbage text in wrong languages is blocked; music lyrics misinterpreted as speech are suppressed via cross-chunk dedup and loop detection
• Encryption — saved files are verified to not contain plaintext; decryption roundtrips are confirmed
• Session cache — write-behind cache verified: files written to disk, data survives simulated crashes, save/load benchmarks under 50ms per flush
• Image PDF OCR — scanned PDFs are extracted via Apple Vision and indexed for search
• Restart reliability — the app can be stopped and restarted multiple times without losing functionality

Verified claims:

Use Case: Interview Preparation

Before the interview

1. Create a playbook with your resume, job description, and prepared Q&As
2. Set it as active
3. Select the Interview prompt

During the interview

1. Click Start Listening before the call
2. The interviewer's voice is transcribed automatically
3. After each question, click Ask AI — the response draws from your resume and previous answers
4. If the topic changes completely, click Clear to start fresh
5. For different question types, switch modes or use quick actions (Answer, Recap, Follow Up)

After the interview

1. Click End Session to see performance metrics
2. Export the session for review

Use Case: Meeting Notes

1. Select the Meeting prompt
2. Click Start Listening when the meeting begins
3. Click Ask AI at any point for a running summary with action items
4. Type: "What did the team decide about the deadline?"
5. Export the session → share structured notes with your team

Use Case: Lecture Companion

Watching a YouTube tutorial, online course, or recorded lecture?

1. Select Explain or General prompt
2. Click Start Listening
3. Type questions as the lecture plays:
   • "Explain that in simpler terms"
   • "Summarize the last 5 minutes"
   • "Give me a code example of that concept"
4. Export as study notes when done

Use Case: Podcast Summarizer

1. Select the Summarize prompt
2. Click Start Listening while the podcast plays
3. Click Ask AI periodically for running summaries
4. Type: "What are the 3 main takeaways so far?"
5. Export as Markdown → structured podcast notes