Source: raw/1_000_days_of_Duolingo_didn_t_work._So_he_built_his_own_app_with_Codex..md

Creator: Claire Vo, How I AI, with William Imoh (captioned “William Emo”), founder of Hackmamba · URL: https://www.youtube.com/watch?v=zBROTIiHusc · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-09). Sponsors: WorkOS, Vanta.

Imoh moved to Norway, kept a 1,000-day Duolingo streak, passed the Norwegian language exam after a one-week grammar course, and still could not hold a conversation, because Norwegians switch to English when you struggle. He had Codex build a native iOS (Swift) and Android (Kotlin) app, Muro, where a voice on GPT Live refuses to speak English and corrects him until he says it right. The second half covers how he runs Hackmamba, a company of about 25 people where he is the only founder: one Codex project per area of the business, a Grok Bot team named after Lord of the Rings characters, and a short list of skills.

Key Takeaways

  • Pick the language for the product, not for yourself. He had never written Swift: “the language I know doesn’t matter anymore. I just choose the best possible language for the tech that I’m building.” Vo contrasts her venture-funded startup of ten years ago, which picked its language by who it could hire.
  • His opening prompt was a product brief. It began “I want a very aesthetically pleasing application”, then:
    • named the models: GPT Live 1 for the voice (“I don’t think I’ll be able to build this without GPT Live 1”) and GPT-6 Luna for other AI tasks;
    • described the interaction: a pulsing orb you talk to, familiar like a phone call;
    • gave style references: “feels like fall, like autumn… use fall colors”, “follow design best practices like you were senior designer at Apple”;
    • set a minimalism rule: “No element on the screen should exist unless it absolutely has to”, and anything unnecessary is removed or put behind a toggle;
    • asked Codex to review the design once it was done. He writes these as epics and user stories, a habit from four years as a product manager.
  • The first build came back in dark mode. His correction: “Not everything belongs in dark mode… light warm colors. Explore gradients and… use liquid glass as much as possible.”
  • Vo’s point: taste is vocabulary. “What you need is design language and references in the world that you can point to.” You need to know what liquid glass, minimalism and a gradient are, and be able to write a user journey any engineer, human or AI, can execute. Her exercise for building taste: describe in words what makes brand A different from brand B.
  • What the app does. It opens with “hi” and gauges your level, corrects tense mid-conversation, and refuses to switch (“I understand but I’m going to speak Norwegian”). The screen shows only speech and a caption. It tracks words you have learned, has flashcards, and offers themed conversations, including a random current topic it researches on the web. When he showed it to others he found that similar apps differed mainly in feel: many were dark-mode and feature-heavy.
  • Cost (his figures). About 5 per 30 minutes of conversation. He open-sourced the app and put 100 was left at recording. Vo: still cheaper than a one-on-one tutor, and available at midnight.
  • GPT Live, per Vo. OpenAI’s real-time voice model; good at real-time use and interruptions, and “really good at… voice control of apps because it’s so fast”. She calls it “totally slept on”. See OpenAI Realtime API.
  • The business runs on a “100x” mandate. From Q3 planning: grow revenue goals 100x without a large headcount increase, using AI as much as possible “without using AI to generate content or seed any of our creative work”. Step one was a Notion list of everything he does: customer success, inbound sales (11 to 15 deals in parallel), the product team, some accounting, and hiring interviews.
  • Tool split.
    • A chief-of-staff hire quit before starting, so he turned to agents.
    • His Claude Cowork account was banned for a reason he does not know, so he moved to ChatGPT.
    • One ChatGPT/Codex project per area: operations (invoicing, campaigns, requests that arrive on Discord), growth, sales, the website, and his podcast (interview briefs).
    • Codex is reserved for complex work, such as the Swift app, where he wants the frontier model.
    • Grok Bot “gave me colleagues whereas with ChatGPT… I got folders.”
  • The Fellowship. He first created a chief of staff named Gandalf, then asked Gandalf to create the rest: Legolas (marketing), an elf the captions render as “Aon” (engineering), Gimli (accounts), Aragorn (sales and recruiting), Sam (customer success) and Frodo (product). Vo adds the bot she says he still needs: “an eye of Sauron that looks across everything for things going wrong.”
  • Access policy: everything except two things. He connected Google Workspace, Linear, Notion, his calendar, Pipedrive and QuickBooks (“everything that I had access to”), and turned on computer use and browser use in Codex with “allow always”, so it can use his computer when there is no MCP connection. The two exceptions: no write access to the bank, and no customer data. He calls this counterintuitive.
  • Systems of record plus an AI interface. His belief: “we’re going to be left with systems of record like a custom database that is designed around a problem… and an AI interface.” His internal tool Three Pipes holds people and hiring records. After a candidate call he tells the bot “I just spoke to this person, get them into three pipes”; it reads his email, calendar and meeting notes, fills the record, summarises, scores the candidate and ranks them (“you’ve spoken to 15 people. This person ranks number seven”). No skill was needed for this.
  • One routine he names: an operator brief every Sunday night with what to start, stop and keep doing, plus the three highest-impact activities for the week.
  • Fewer skills over time. As models improved he dropped most of his skills. The ones left: growth-marketing strategy and content strategy (both “based off what we know at Hackmamba”) and a content-deslop skill.
  • When an agent stalls. Restart it. Then get more specific: “do this first, test this first, and also review what you’ve done to be sure this is what I ask for.” Recently his Grok Bot drives his Cursor agent and reviews and improves what it produced. Failing that, switch agents.
  • The risk he names. “I’m so dependent on agents now that I fear for myself… What if these agents go to zero tomorrow?”

Try It

  1. Write your next build prompt as a product brief: the models to use, the core interaction, style references in named design vocabulary, a minimalism rule (“no element unless it has to exist”), and a closing request to review the design.
  2. Build the vocabulary first. Each week, write a paragraph on what makes one site or app different from a competitor (Vo’s exercise).
  3. For language practice, set up a voice tutor on a real-time voice model that refuses your native language and keeps correcting until you get it right. Budget his 5 per 30 minutes. The app is open source, but the transcript gives no repository link.
  4. Add a Sunday routine: start, stop, keep, and the week’s three highest-impact activities.
  5. Write your agent access policy as a short list of exclusions and check whether his two (bank write access, customer data) cover your real risks.
  6. Replace one skill with a system of record and a plain request, the Three Pipes pattern: a purpose-built table the agent fills from email, calendar and notes.
  7. Have one agent review another’s output before you look at it, as his Grok Bot does for Cursor.

Open Questions

  • The app’s name is captioned as “Mural Chat”, “Muro” and “Euro”; “Muro” is used here. The guest’s surname is captioned “Emo”; Hackmamba’s founder is normally written William Imoh.
  • The open-source repository URL is not in the transcript.
  • He does not know why his Claude account was banned.
  • The engineering elf’s name is unclear in the captions.
  • His 5 per 30 minutes is a recollection on air, not a measured cost.