Interview mode

The interview copilot.

It hears the interviewer, drafts the answer as the question lands, and puts it where only you can see it. Short enough to use mid-sentence.

Transcript
InterviewerDesign a URL shortener. Assume one of the links goes viral.
YouLet me scope it first — how many new links a month are we expecting?
InterviewerSay a hundred million, mostly re
Q: URL shortener that survives a viral link
Design
  • Requirements: 100M links/month, 10:1 reads, p99 < 100 ms
  • API: POST /shorten · GET /{code} → 302
  • Data model: code → long URL, owner, expiry
  • Core flow: hash, collision check, stateless redirect service
  • Storage: key-value store, codes as keys
  • Scaling: cache hot codes, shard by code prefix
  • Failure modes: hot key on viral link → CDN + in-memory cache
Say

I’ll assume a hundred million new links a month, reads ten to one over writes, and a p99 under 100 milliseconds on the redirect. The headline is a stateless redirect service over a key-value store, with an in-memory cache and CDN edge caching for the hot codes.

Trade-offs
  • Key-value over relational because lookups are by code only (accepting no rich queries)
  • 302 over 301 because analytics need every hit (accepting one extra hop)
⌘⇧A Ask⌘⇧S Screenshot⌘⇧M Mic⌘⇧\ Hide⌘ on macOS · Ctrl on Windows · global, even while the call has focus
Inside a session

What happens between the question and your first word.

Both sides of the call, transcribed live

Your microphone and the system audio are captured as two channels, so the interviewer and you appear as separate, labelled bubbles. Each bubble updates in place as the words arrive and locks when the sentence ends. Live transcription runs in 45+ languages, including bilingual pairs.

InterviewerWalk me through how you’d shard this table.
YouI’d start from the access pattern —
InterviewerAnd if one tenant is ten times the

Scan to glance at. Say to read aloud.

Every answer has a fixed shape so you never have to read to find the point. Scan is 3–5 bullets of eight words or fewer, the direct answer first. Say is the same answer as 50–90 words of first-person prose. Coding questions get Say · Code · Explain; system design gets Design · Say · Trade-offs; behavioral questions get four STAR beats and a 90–130-word story.

Concept · MLScanSay
CodingSayCodeExplain
System designDesignSayTrade-offs
BehavioralScan (STAR)Say

Go Deeper when they follow up

One click on any answer adds 250–350 words that add the two insights that matter most, and never restates the original. In a system-design round, each Go Deeper expands the next stage the interviewer will ask about, in the order they ask it: requirements, API, data model, components, storage, scaling, failure modes. ML designs follow their own ladder from data and labels through serving and evaluation.

  1. Requirements & numbers
  2. API
  3. Data model ← Go Deeper
  4. Core components
  5. Storage
  6. Scaling
  7. Failure modes & consistency
  8. Observability & rollout

Take a screenshot for written questions

Cmd/Ctrl+Shift+S captures the screen. A coding problem in the editor comes back as a one-line approach to say, a full runnable solution in the language on screen, and a two-bullet explanation with the complexity. A diagram or a design prompt gets the design walkthrough. Screenshots are deleted when the session ends, never kept long term.

Code
def two_sum(nums, target):
    # one pass: value -> index
    seen = {}
    for i, n in enumerate(nums):
        if target - n in seen:
            return [seen[target - n], i]
        seen[n] = i
Complexity: O(n) time, O(n) space

Your résumé, the job description, your notes

Upload files once and they become reference material for every answer. Set the company, role, and round type when you start a session; they never leak from a previous one. Answers are calibrated to your seniority, draw on your résumé and documents, and are told to stick to them. AI can still get details wrong, so check what it says about your own work before you use it.

resume.pdfjob-description.txtsystem-design-notes.md
CompanyAcme Robotics
RoleSenior ML Engineer
RoundSystem design

History, a summary, and a choice of model

Sessions are saved to your local history with their transcript. When you end it, you get a recap of what came up, a score out of 10 with one-line call-outs on clarity, depth, and structure, and what to tighten next time. Free answers with Claude Haiku; Pro lets you pick any of the six models.

  • Claude Haiku 5.5 (fast, default)
  • Claude Sonnet 5.5 (sharper, Pro)
  • Gemini 3.5 Flash-Lite (fast, Pro)
  • Gemini 3.8 Flash (sharper, Pro)
  • ChatGPT Luna (fast, Pro)
  • ChatGPT Sol (sharper, Pro)
Screen share

The overlay is excluded from screen capture.

It uses the operating system’s own screen-share exclusion, so it never appears in a shared screen or a recording. It stays out of the taskbar, can be dimmed to a whisper, and can pass clicks straight through.

Three modes, in the app or the browser

One app for the interview, the rehearsal, and the job after.

Keyboard

Every action is a shortcut.

Cmd on macOS, Ctrl on Windows. The four global ones fire even while Zoom has focus.

Global work even while the call has focus

  • ⌘/Ctrl⇧AAsk from voice
  • ⌘/Ctrl⇧SScreenshot
  • ⌘/Ctrl⇧MToggle mic
  • ⌘/Ctrl⇧\Hide / show app

In-app when the overlay is focused

  • ⌘/Ctrl⇧RReset feed
  • ⌘/Ctrl⇧TClear transcript
  • ⌘/Ctrl⇧↑Scroll up
  • ⌘/Ctrl⇧↓Scroll down
  • ⌘/Ctrl⇧=Bigger font
  • ⌘/Ctrl⇧-Smaller font
  • ⌘/Ctrl⇧CCopy response
  • ⌘/Ctrl⇧]More opaque
  • ⌘/Ctrl⇧[More transparent
  • ⌘/Ctrl⇧GSee-through (clicks pass through)
Interview mode

Questions about the copilot.

Yes, with the browser version: open sokrates.cloud/companion in Safari, Chrome or Edge on a phone, tablet or computer, with nothing to install. On its own it listens through that device's microphone, so the interviewer has to come out of the speakers (no headphones). It is an ordinary browser tab, so it shows up if you share the screen it is on; only the desktop app is excluded from screen capture. Paired with the desktop app, a phone shows the live answers under your webcam and its buttons trigger Answer now, Screenshot, Go Deeper and Hide on the computer; headphones are fine that way. Add to Home Screen installs it as a plain "Sokrates" icon.
No. With the mic on, Sokrates listens to the interviewer channel and drafts an answer as each question lands. Cmd/Ctrl+Shift+A asks from voice on demand, and Cmd/Ctrl+Shift+S takes a screenshot when the question is written rather than spoken.
Short by design. The Scan section is 3–5 bullets of eight words or fewer; the Say section is 50–90 words of first-person prose you can read aloud. Coding answers add a code block and a one-line complexity note. Go Deeper adds 250–350 words when you want them, and never restates the original.
The first answer lays out the design stage by stage (requirements, API, data model, components, storage, scaling, failure modes; or the ML equivalent). Each Go Deeper expands the next stage the interviewer will ask about, so the conversation is walked in order rather than dumped at once.
Your résumé, the job description, any documents you upload, plus the company and role you set for the session. Answers are calibrated to that seniority, draw on your résumé and documents, and are told to stick to them. AI can still get details wrong, so check what it says about your own work before you use it.
Yes. The chat box takes a typed question in any mode, with the recent transcript as context.
A short wrap-up: a recap of the questions that came up, a score out of 10 with one-line call-outs on clarity, depth, and structure, and what to tighten next time. The session is saved to your history.

Ready before the next question.

Free to download. Ten minutes of live transcription a day on the free plan; Pro removes the caps.

Free plan: 10 minutes a day, no card · Apple Silicon · Windows 10/11 x64