ExpertsFMHow this founder built a better Duolingo with Codex | William Imoh
How I AI · November 9, 2026
5 Selected Lessons:
framework · 2 reasoning_pattern · 1 heuristic · 1 case · 1
Detailed, reference-rich prompting to encode design taste into an agent build
Source
How I AI — How this founder built a better Duolingo with Codex | William Imoh (November 9, 2026) · at 7:21 ▸ · at 11:07 ▸
When this applies
when directing any AI agent to produce a creative or design-driven output without a human designer in the loop
The takeaway
The richer and more specific your reference set and vocabulary for what you want, the better an AI agent can execute your intent without additional human expertise.
Why it works
Specific design vocabulary (minimalism, liquid glass, warm fall colors, 'senior Apple designer') gives the model concrete stylistic anchors it can execute against, substituting for a human designer's taste.
One level up
In AI-mediated creation, the bottleneck shifts from technical execution skill to articulation skill — the ability to describe desired outcomes precisely.
Where it stops holding
This stops working when the person giving the prompt lacks any design/product vocabulary to reference, producing vague prompts the agent can't act on well.
Detail
procedure:
- Open with an aesthetic mandate ('I want a very aesthetically pleasing application')
- Specify the exact tech/model to use (e.g., GPT Live One)
- Describe the core interaction experience (e.g., 'pulsating orb like a phone call')
- Specify stylistic references (warm, fall colors, minimalist, liquid glass, 'senior designer at Apple')
- State a feature-minimalism rule (no element unless absolutely necessary, else toggle it)
- Have the agent review its own output after producing it
benchmarks: {}
rationale: Precise design language and product/user-story framing (from William's PM background) let the agent produce a polished result on the first detailed pass, refined further through iteration.
In their words
Use fall fall colors, and again, it should be really beautiful. Follow design best practices like you are a senior designer at Apple.
Siloed 'projects' per business function to manage AI-assisted context
Source
How I AI — How this founder built a better Duolingo with Codex | William Imoh (November 9, 2026) · at 24:54 ▸ · at 28:11 ▸
When this applies
when one person must juggle many distinct workstreams with AI assistance and needs to avoid context bleed
The takeaway
Partition AI context boundaries along the same lines you would partition organizational responsibilities, so each AI workspace mirrors a real functional silo.
Why it works
Dedicating a separate AI 'project'/context per function prevents cross-contamination of unrelated information and lets each project accumulate relevant memory/routines specific to that domain.
One level up
Structure AI tool usage to mirror organizational structure, not just task-by-task, for sustained operational leverage.
Where it stops holding
This stops working when functions are too interdependent and siloing loses important cross-functional context that a unified view would catch.
Detail
procedure:
- Create one project per critical business function (operations, growth, sales, content/podcast)
- Route related inbound messages (e.g., Discord) into the matching project
- Set up recurring routines per project (e.g., weekly 'operator brief' summarizing start/stop/keep doing and top 3 impact activities)
benchmarks: {}
rationale: This structure let a solo founder manage customer success, sales, product, accounting, and HR simultaneously without losing track of context.
In their words
I set up projects for everything basically, because I just wanted a siloed way to manage individual things I do.
Treat any skill gap as a solvable problem via AI-assisted iteration
Source
How I AI — How this founder built a better Duolingo with Codex | William Imoh (November 9, 2026) · at 0:00 ▸ · at 20:02 ▸
When this applies
when facing any skill you lack, if you believe it's a solvable problem and can iterate with feedback, you can acquire it quickly using AI as tutor/collaborator
The takeaway
A growth-mindset belief that a domain is 'solvable' lets you convert AI's patience and correction into real competence, regardless of starting skill level.
Why it works
AI removes the traditional barrier of needing pre-existing expertise by providing infinitely patient, non-judgmental feedback loops, so confidence plus iteration substitutes for prior knowledge.
One level up
In an age of capable AI assistants, confidence and willingness to ask questions and be corrected is a higher-leverage skill than pre-existing domain expertise.
Where it stops holding
This stops working when the person is too ego-invested to accept correction, or when the domain requires physical/embodied practice AI cannot simulate.
Detail
approach: Believe the goal is solvable; be low-ego about mistakes; use AI as an infinitely patient tutor; ask questions and get feedback to iterate toward competence.
signals:
- confidence framing a new domain as learnable
- willingness to be corrected repeatedly without frustration
- using AI across multiple unfamiliar domains (coding, language)
rationale: AI's patience and correction ability lets confident, iterative learners close skill gaps that previously required years of study or expert mentors.
In their words
The meta skill that you have, which is like a growth mindset, which is I can learn Norwegian. Like, I can learn something. If you just have this confidence that this thing is a solvable problem, learning a language is a solvable problem, building an app is a solvable problem, and if you can just ask the right questions and get the right feedback, you can get there.
Choose the best tool/language for the experience, not the one you know
Source
How I AI — How this founder built a better Duolingo with Codex | William Imoh (November 9, 2026) · at 5:08 ▸ · at 7:21 ▸
When this applies
when choosing any implementation tool or medium, now that AI agents can execute in tools you personally don't know
The takeaway
When execution capability is no longer scarce or operator-specific, choose means based on end-outcome quality, not on who/what already knows how to do it.
Why it works
AI agents can write code in any language competently, so the constraint of 'what do we/our team know' is removed, freeing the choice to be driven purely by fitness for the end-user experience.
One level up
Remove self-imposed constraints once the resource that created them (human capability) becomes commoditized by a new tool.
Where it stops holding
This stops working when the AI agent is not actually competent in the chosen language/tool, or when maintainability requires ongoing human expertise the team lacks.
Detail
rule: Pick the best language/tool for the desired user experience; let AI agents handle execution regardless of your own fluency.
threshold: Use native premium frameworks (e.g., Swift) when a 'premium feel' is the goal, even with zero personal experience in that language.
rationale: AI coding agents decouple tool choice from human skill inventory, so optimizing for UX quality beats optimizing for team familiarity.
In their words
I guess when we have AI agents building stuff, the language doesn't matter. Like the language I know doesn't matter anymore. I just choose the best possible language for the type that I'm building.
Dark mode rejection and iterative design refinement
Source
How I AI — How this founder built a better Duolingo with Codex | William Imoh (November 9, 2026) · at 11:07 ▸
When this applies
when a first AI output misses the intended tone or style
The takeaway
Treat the first AI output as a draft to be critiqued with increasingly specific corrective language, not as a final product to accept or reject wholesale.
Why it works
Giving corrective, specific feedback ('not everything belongs in dark mode,' use light warm colors, gradients, liquid glass) let the agent converge quickly on the desired aesthetic.
One level up
Iteration speed with specific feedback beats trying to get a perfect prompt on the first try.
Where it stops holding
This stops working when the feedback given is vague rather than specific, leaving the agent without clear direction to improve.
Detail
situation: Agent produced a dark-mode design that didn't match the desired 'fall/autumn' warm aesthetic.
decision: He explicitly rejected dark mode and specified light mode, warm colors, gradients, and liquid glass.
outcome: The redesign succeeded and the app's distinctive minimalist look (vs. competitors' dark, feature-heavy versions) became a differentiator that drove positive public reception.
lesson: Specific negative feedback ('not this') paired with specific positive direction ('this instead') is an efficient way to correct AI-generated creative work.
In their words
The first design he produced was shit. It was in dark mode. It was dark mode. I was like, no, this isn't gonna work.