Context
William Imoh is the founder of Hackmamba, a developer marketing company, and the solo builder of Mural, a real-time voice app for practicing Norwegian. He has a chemical engineering degree, taught himself web development, and spent four years as a product manager, but he had never written Swift. In this How I AI episode, host Claire Vo has him walk through two things: how he got Codex to build Mural as a native iOS and Android app, and how he runs a team of about 25 people with agents. For PMs, the episode is a clear case of product and design skills becoming the main input to AI building, plus a candid look at what giving agents broad access looks like in practice.
The Big Idea
When an AI agent writes the code, the scarce skill is no longer knowing the programming language. It is describing the experience you want with enough precision, in design and product language, that any engineer, human or AI, can build it.
William got a polished native app from a long, specific prompt about look, feel and user journey, not from Swift knowledge. A year earlier the same idea, built with older tools, came last at a San Francisco hackathon.
Key Insights
Streaks are not fluency
William's problem came from a real gap between practice metrics and the actual skill:
- He kept a 1,000-day streak on Duolingo, always chasing the streak, and still could not hold a basic conversation in Norwegian.
- Norwegians are polite: when he tried weak Norwegian, they switched to English, so he never got the feedback loop of "did I say it right?"
- He took a one-week grammar course, passed the Norwegian language exam, and deleted Duolingo on the way out of the exam hall because it had not helped him speak.
- What he wanted instead was something that would not judge him, would correct him, and would feel like being dropped into a Norwegian village with no option but to speak Norwegian.
The PM lesson is that an engagement metric like a streak can look healthy while the user's real job, speaking with people, goes undone.
Pick the stack for the user
William chose native Swift for iOS and Kotlin for Android because he wanted a premium feel and access to the best iOS features, even though he had never written a line of Swift. His reasoning was that with AI agents building, the language he personally knows no longer matters, so he picks the best language for what he is building.
Claire contrasted this with her own venture-funded startup over 10 years ago, where the team chose its language based on which people they knew could write it. With Codex "on staff," that constraint goes away, and the technical choice can follow the user experience instead of the hiring pool.
Design vocabulary is the new skill
Claire's strongest point was that what William had, and what many people have not invested in, is design and product language:
- Taste in practice means having design language and references you can point to, such as knowing what liquid glass, minimalism and a gradient are.
- The richer your set of design references, the better the agent output.
- On the product side, the skill is articulating a user journey so clearly that any engineer, human or AI, can execute it.
- Her way to build this: practice describing websites and apps in words, saying what makes brand A different from brand B, or why one app is better than another.
William's background helped here. As a former PM he works in epics and user stories because they stay tied to user problems, and he said that gave him his best results building with AI.
One prompt set the bar
William's first Codex prompt was long and very specific. It included:
- An opening line asking for a very aesthetically pleasing iOS app for learning Norwegian, which he called his version of "make no mistakes."
- The models: GPT-Live-1 for the real-time voice conversation and GPT-6 Luna for any other AI tasks.
- The interaction: a pulsating orb you talk to that feels like a phone call, a familiar interface.
- The style: autumn colors, "like you are a senior designer at Apple," and strict minimalism.
- A minimalism rule: no element on screen unless it has to be there, and anything unnecessary is removed or put behind a toggle.
- A self-check: review the design after finishing it.
The first design still missed. It came back in dark mode, which he rejected, then redirected to light mode, warm colors, gradients and as much liquid glass as possible.
Experience beat feature count
William did not plan to share Mural, because he did not want the burden of maintaining it or managing user expectations. When he posted it and people loved it, he compared it with similar apps others had built. The difference was mostly experience: how it felt to look at. Many of the others were in dark mode and packed with features. Mural's core was deliberately narrow, only speech and a caption, since he believed a caption is the only way a beginner can follow a language they are hearing for the first time.
A tutor that refuses English
The live demo showed the product decisions that make Mural work as immersion:
- It opens with "hi" and then works out the learner's level from how they respond.
- It corrects mistakes as they happen, for example fixing the wrong form of the word for "meeting."
- When William repeatedly asked it to speak English, it refused and kept asking about the meeting in Norwegian.
- It keeps correcting until the learner says the phrase properly.
- Later additions include a tracker of learned words and how well each is going, flashcards, and themed conversations, including a random theme where it searches the internet for a current topic.
The refusal to switch languages is the feature. It deliberately removes the polite escape hatch that stopped William from learning in real life.
Real-time voice has a cost
Claire called GPT-Live, OpenAI's real-time voice model, "totally slept on." It handles real-time conversation and interruptions well, which suits language practice, and it is fast enough for voice control of apps. On cost, William was direct:
- A 30-minute Mural session costs about $3 to $5.
- He made Mural open source so anyone can run it.
- He put $1,000 on his OpenAI budget for people to use, and about $100 was left.
Claire's counterpoint was that this is still cheaper than a one-on-one tutor, and a tutor cannot be called at midnight when you feel like practicing.
Agents gave him colleagues
At Hackmamba, William's work spans customer success, most inbound sales (about 11 to 15 deals in parallel), the product team, accounting, HR and interviews, for a team of around 25. His attempt to hire a chief of staff failed when the hire quit before starting. His agent setup grew from that gap:
- He tried Claude first, but his account was banned for reasons he does not know, so he moved to ChatGPT and Codex.
- In Codex he made a project for each critical area he works on directly, such as Hackmamba operations, growth, sales, the website, and his podcast.
- He reserves Codex for complex work, like the Swift code for Mural, where he wanted a frontier model.
- Grok Bot felt different. In his words, ChatGPT gave him Codex and folders, while Grok Bot gave him colleagues.
Broad access, two hard limits
William deliberately ignored the usual caution about what agents can read and write. His agents have access to nearly everything: Google Workspace, Linear, Notion, his calendar, Pipedrive, QuickBooks, and roughly 15 or 16 browser tabs of tools. In Codex he turned on computer use and browser use, so if a connector cannot find something, the agent uses his computer, and he chooses "allow always" over "allow once."
His two limits are that agents have no write access to the bank and no access to customer data. The payoff is that he no longer feeds agents information. He can type a rough thought, typos included, and the agent finds the context itself.
Systems of record plus AI
William believes that in the end companies will be left with systems of record, meaning custom databases designed around one problem, plus an AI interface on top. So he builds internal tools that serve as those systems of record.
His hiring tool shows how it works. After talking to a candidate, he tells his Grok Bot to add that person to the people and HR tool. The agent checks his email, calendar and meeting notes, adds the record, writes a summary, and scores the candidate against everyone else, for example telling him a person ranks seventh out of the 15 he has spoken to. He did not need a skill for this, and he can switch straight to sales afterward.
Fewer skills as models improve
William has built many skills over time, but the models got good enough that he no longer needs most of them. The few he keeps are specific:
- A growth marketing strategy skill based on what Hackmamba knows.
- A content strategy skill, also based on company knowledge.
- A content "deslop" skill, which Claire called critical.
This matches his "100X" rule, explained below, that AI should not generate the company's creative content.
Mental Models & Frameworks
The 100X mandate
In June, while planning Q3, William set an internal mandate called 100X: aim for 100 times Hackmamba's current revenue goals without drastically increasing headcount. The mandate had two parts:
- Use AI as much as possible across the business.
- Do not use AI to generate content or hand any creative work over to AI tools.
His first step was simple: list everything he personally does in Notion. That inventory became the map of which roles agents should take on.
The agent fellowship
William set up Grok Bot as a team themed on The Lord of the Rings, matching each character to a role:
- Gandalf: chief of staff, created first, then asked to create the rest of the team.
- Legolas: marketing.
- Gimli: accounts, described as the one with the coins.
- Aragorn: sales, since he recruits everyone to the fellowship.
- Sam: customer success, since he always plays support.
- Frodo: product management, the product owner, since Frodo is in charge of getting the ring.
- An engineering bot also sits in the fellowship.
Claire suggested one more role: an "eye of Sauron" agent that looks across everything for things going wrong. The model works because each agent has a clear role, so a request goes to a colleague with a job rather than into a generic chat.
One project per workstream
William's Codex setup gives every critical workstream its own project, so context and tasks stay separate:
- Operations: invoicing, campaigns, and messages that come in on Discord.
- Growth: includes a routine that sends him an operator brief every Sunday night.
- Boki: Hackmamba's internal tool for customers, covering content from ideation to distribution.
- Sales, website and podcast: the podcast project, for Everything Outside Code, handles interview briefs.
The Sunday operator brief
A recurring routine in the growth project sends William a brief every Sunday night. It answers:
- What should we start doing?
- What should we stop doing?
- What should we keep doing?
- What are the three highest-impact activities for this week?
It is a lightweight weekly review that an agent runs for a solo founder, so the week's priorities are set before Monday.
The stuck-agent escalation ladder
When an agent will not do what he wants, William climbs a short ladder:
- Restart the agent and try again.
- Get more specific: do this first, test this first, then review what you did to check it matches what was asked.
- Have a second agent review the first. He now has his Grok Bot drive his Cursor agent, then review and improve its less accurate work.
- Switch to a different agent and wait for the original tool to improve.
He said he can be this specific because he already knows how to do the task. He just does not want to do the low-level work.
Trade-offs & Nuance
Open access versus risk
William admitted that being skeptical about what AI can read and write is the common view, and he chose to set it aside. His position is not "no limits." It is broad access with two firm exceptions: no write access to the bank and no customer data. The trade-off is speed and low friction for him, against a larger blast radius if an agent misbehaves. A PM copying this setup should decide which limits match their own risk.
Dependency is the real fear
William said agents do not frustrate him. What scares him is how dependent he has become. If agents "go to zero tomorrow," he does not know how he would manage 15 sales opportunities a week alongside calls and hiring. His fear is not being replaced, it is relying on tools he does not control. Leverage and fragility grow together.
Cost versus access
Mural's voice sessions cost real money, about $3 to $5 per 30 minutes, and William is paying for strangers to use it as a hobby. That works for a fun open-source project. A paid product would need a different answer. Claire's framing that it is still cheaper and more available than a human tutor is the comparison a real business case would rest on.
Building for one user
William built Mural only for himself and did not want the burden of a business. That focus is likely why the product is so narrow and opinionated. It also means he has to balance public interest against his choice not to take on users' expectations.
Practical Application
Write a design-spec prompt
Before asking an agent to build a UI, write the prompt the way William did:
- Name the feeling you want, such as "feels like autumn," with concrete colors.
- Give a reference persona, such as "a senior designer at Apple."
- Set a removal rule: no element on screen unless it must be there, and anything optional goes behind a toggle.
- Name the interaction metaphor, such as an orb that feels like a phone call.
- Ask the agent to review its own design when it finishes.
Expect the first result to miss, as William's dark-mode first draft did, and plan a refinement pass.
Practice describing products aloud
Claire's way to build taste is to describe websites and apps in words. Once a week, pick two competing products in your space and write a few sentences on what makes each one feel different, using specific design terms. The goal is a richer vocabulary you can reuse in prompts and in briefs to designers.
Audit metrics against the job
William's 1,000-day Duolingo streak did not lead to speaking Norwegian. Pick one engagement metric your team celebrates, such as streaks, daily active use or sessions, and check it against the outcome users actually hire the product for. If the two can diverge, add a measure of the real outcome.
Set up a weekly operator brief
Create an agent routine that sends you a brief every Sunday night with what to start, stop and keep doing, plus your three highest-impact activities for the week. Point it at the tools where your work lives, such as calendar, project tracker and notes, so it reasons from real activity.
List your work before automating
Do what William did before his 100X mandate: write down every task you personally do in a single doc. Group the tasks into roles, such as sales, product, HR and customer success, and only then decide which roles an agent should take, with what access, and which work stays human.
Questions to Consider
- Which engagement metric on your product, like a streak or daily usage count, could look healthy while users still fail at the outcome they came for, the way a 1,000-day Duolingo streak left William unable to speak Norwegian?
- If an AI agent can write any programming language, which technical choices on your team are still being made around who knows what, rather than around the best user experience?
- Could you describe your product's intended look and feel precisely enough, in design terms like minimalism, gradients or liquid glass, that an AI agent could build it without a mockup?
- Where would you draw the access line for your own agents, given that William gives his agents nearly everything except write access to the bank and access to customer data?
- If the agents your team relies on stopped working tomorrow, which workflows would break first, and what is your fallback?
Bottom Line
William Imoh built a native voice tutor in a language he had never written because he could describe exactly what he wanted, in design and product terms, and Codex handled the code. For PMs, the durable advantage is a precise vocabulary for experience and user journeys, paired with clear decisions about what agents can access and what stays human.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| Mural (open source) | Code repository | William made Mural open source so anyone can run their own copy of the Norwegian voice tutor. |
| Everything Outside Code | Podcast | William's own podcast, which he manages through a dedicated Codex project that handles interview briefs. |
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Codex | OpenAI's coding agent, with projects, routines, computer use and browser use | William used it to build Mural in Swift and Kotlin and organizes each business workstream as a Codex project. |
| GPT-Live | OpenAI's real-time voice model | It powers Mural's spoken conversations; William said he could not have built the app without it. |
| GPT-6 Luna | An OpenAI model | William's prompt specified it for every AI task in Mural other than the live voice conversation. |
| Grok Bot | A personal and work AI agent | William runs his Lord of the Rings "fellowship" of role-based agents on it and has it review his Cursor agent's work. |
| Cursor | An AI coding tool | William now has his Grok Bot drive the Cursor agent, then review and improve its output. |
| Boki | Hackmamba's internal growth tool | Hackmamba uses it with customers for content from ideation to distribution, and it has its own Codex project. |
| Notion | Docs and workspace tool | William listed every task he does in Notion as the first step of his 100X mandate, and his agents can access it. |
| Linear, Pipedrive, QuickBooks | Project tracking, sales CRM and accounting tools | These are among the business tools William gave his agents full access to. |
People to Follow
William Imoh
William is the founder of Hackmamba, a developer marketing company, and creator of Mural, an open-source real-time Norwegian voice tutor. He is a former product manager of four years with a chemical engineering degree, and he runs a team of about 25 as a solo founder with heavy use of agents. He hosts the podcast Everything Outside Code and offered to talk with developer tool companies about growth.
Experiments to Try
Experiment: Design-vocabulary prompt test
- Hypothesis: a prompt rich in specific design references produces a better first UI draft than a short functional prompt.
- Change: build the same screen twice with an AI coding agent, once with a plain feature description and once with a William-style prompt naming mood, colors, a reference designer persona, a minimalism rule and a self-review step.
- Test: show both drafts to a few teammates without saying which prompt produced which.
- Measure: preference, number of revision rounds needed to reach an acceptable design, and how many unnecessary elements each draft includes.
Experiment: Agent-reviews-agent pass
- Hypothesis: having a second agent review a coding agent's output catches errors earlier than a single agent working alone, as William found with Grok Bot reviewing his Cursor agent.
- Change: for a week of small coding or doc tasks, add a review step where a second agent checks the first agent's work against the original request.
- Test: compare against a similar set of tasks done with one agent only.
- Measure: defects found by a human afterward, number of follow-up prompts, and total time to an accepted result.
Notable Quotes
"I just choose the best possible language for the tech that I'm building." (William Imoh)
"No element on the screen should exist unless it absolutely has to." (William Imoh)
"Really what you need is design language and references in the world that you can point to." (Claire Vo)
"I say it is an infinitely patient, private tutor." (Claire Vo)
"I get scared because I'm so dependent on agents now that I fear for myself." (William Imoh)
