CHAPTERS
- 0:00 – 1:19
Why prompts fail: unrealistic expectations and missing context
X.K. argues that the idea of a single “perfect prompt” is a myth because the model doesn’t know your goals, constraints, or progress. Prompt tweaking is often a symptom of insufficient context plus mismatched expectations about what AI can do from a short instruction.
- •One-shot prompting is like hiring someone and expecting results in 5 seconds with no onboarding
- •Prompt failures come from both model limits and expectation management
- •Vibe-coding and quick prompting often ignore project state, constraints, and intent
- •The real fix is not better wording, but richer context over time
- 1:19 – 5:23
Context engineering as a product behavior: accumulate materials, compound results
Rather than treating AI as a one-off chat tool, Kuse pushes users to build an evolving context base. Over time the system “knows you more,” reducing the need for complex prompts and improving outputs via a compounding loop.
- •Two dimensions: technical context + managing user expectations
- •Encourage users to store and accumulate files, notes, and artifacts in one place
- •AI improves as it reuses prior context (patience + compounding)
- •Provide non-verbal ways to express intent beyond plain text prompts
- 5:23 – 7:23
The real story behind “$10M ARR in 60 days”: long build-up before the spike
X.K. clarifies that the headline compresses a much longer journey—Kuse was built quietly starting early 2024. The “60 days” reflects an inflection after sustained product development, community relationships, and groundwork.
- •Product development started well before the viral growth window
- •Built attention and connections in local communities (HK/Taiwan)
- •Avoided big marketing campaigns early, staying under the radar
- •The spike came after accumulated distribution + product readiness
- 7:23 – 9:07
Pivot narrative: from design agent to horizontal knowledge/document AI
Kuse began as a design agent, which explains the infinite canvas UI. Real usage showed people valued uploading and analyzing documents far more than generating designs, prompting a shift toward knowledge-work and document understanding.
- •Initial vision: design agent that converts requirements into designs/posters
- •Users primarily used it as a document/file analysis workspace
- •Image models and design outputs weren’t meeting expectations at the time
- •Team doubled down on the knowledge base + document workflows
- 9:07 – 11:28
Threads growth playbook: underserved platform + geography + intern ‘content army’
Kuse’s distribution edge came from going all-in on Threads—especially in Taiwan and Hong Kong—where it was growing fast but less crowded. They scaled content via many accounts run by interns, publishing daily use cases that drove awareness and demand.
- •Threads was ignored in the US but strong in HK/Taiwan markets
- •Created hundreds of accounts to post use cases at high volume
- •Threads’ algorithm distributed traffic generously due to less entrenched creator hierarchy
- •Key viral features: exam paper generation and “Formatter” (layout/formatting agent)
- •Also found Instagram effective; X is harder for organic growth without network effects
- 11:28 – 12:20
Threads accounts demo: what content formats worked and why
X.K. shows examples of their many Threads accounts and the kind of posts that performed—practical, repeatable workflows and before/after transformations. They discuss market competition dynamics and why Threads had more “real users” than some assume.
- •Demonstrates Formatter-style posts turning messy text into polished layouts
- •Confirms most visible accounts in the demo are owned/operated by Kuse
- •Highlights less intense competition in HK/Taiwan vs US/China
- •Explains why X has a rigid hierarchy; Threads offers easier discovery for newcomers
- 12:20 – 17:06
Visual context engineering: a marketing term for multi-modal intent expression
X.K. frames “visual context engineering” as a shorthand to communicate their approach: using a 2D canvas to express intent and organize context. The canvas enables spatial relationships among documents, images, and outputs—making reuse and iteration easier than linear chat.
- •Term created to explain the product succinctly (not purely technical)
- •Users can express intent via spatial organization, sketches, grouping, and selection
- •2D workspace makes documents visible and reusable vs buried chat attachments
- •Enables a loop: create → store in library → reuse → refine with less effort
- 17:06 – 20:10
The “mom cooking” analogy: context is personalized understanding + feedback loops
To make context engineering intuitive, X.K. compares it to a mom who knows a child’s preferences and goals and cooks accordingly. The better the context, the better the output, reinforcing a positive loop of trust and usefulness.
- •Context = preferences, goals, constraints, and history
- •Better context yields better “tailored” outputs with less instruction
- •Repeated interactions deepen understanding and improve results
- •Explains why generic prompts produce generic answers
- 20:10 – 22:12
RAG vs fine-tuning vs prompt engineering—and Kuse’s async file processing twist
They discuss common confusion between RAG, fine-tuning, and prompting, then outline Kuse’s approach. Kuse emphasizes RAG and robust document/OCR processing, with async pre-processing so files are ready before users query them.
- •Fine-tuning is heavy; Kuse relies more on RAG and document processing
- •Async ingestion: pre-process files in folders even when users are away
- •“Prepared ingredients on the table” metaphor for faster downstream work
- •Product focus: file management + OCR + retrieval quality over fancy prompting
- 22:12 – 26:26
MVO before MVP: validate model outputs before productizing features
X.K. introduces an internal product principle: build a Minimal Viable Output (MVO) before investing in a Minimal Viable Product (MVP). In AI products, if the model output isn’t reliable, shipping UI and workflows is wasted effort.
- •Traditional flow (PRD → build → launch) breaks when AI output quality is unstable
- •Start with output experiments and stabilization first
- •Only productize once outputs are consistently useful
- •Context techniques are judged by output quality, not architecture elegance
- 26:26 – 31:43
Demo: creating a PRD in Kuse with minimal prompting but heavy context
X.K. walks through Kuse’s workflow: drop files onto the canvas, select relevant materials, and ask for an artifact (like a PRD). The emphasis is on providing background docs, notes, and constraints so a simple prompt can produce strong results.
- •Three-step workflow: drop files → select → ask → get artifact
- •“Source only” option constrains outputs to uploaded materials
- •Supports multiple models (GPT, Claude, Gemini, etc.)
- •Key to quality PRDs: upload project background, progress, decisions, open questions, and meeting notes
- 31:43 – 32:48
Prototype generation and the ‘AI wrapper’ debate: usefulness over complexity
They show generating a simple prototype via webpage generation, noting that the system summarizes context and hands it to a strong model (e.g., Claude). X.K. argues teams shouldn’t overcomplicate solutions just to avoid the “wrapper” label—users only care if it solves the problem.
- •Behind the scenes: summarize/label docs, then delegate generation to a model
- •Claude highlighted as strong for coding/prototyping tasks
- •Positioning: Kuse competes by compounding project context over time
- •Targets non-coders differently than tools like Claude Code/Cursor
- •Use cases extend beyond prototyping (e.g., HR/admin announcement pages)
- 32:48 – 44:43
Founder guidance and philosophy: don’t fear model updates; follow users; think beyond productivity
X.K. advises founders to avoid FOMO and loss aversion around rapid AI progress and instead stay close to users. He frames Kuse long-term as more than a productivity tool—potentially a “playground” that preserves meaning and satisfaction in a world where AI automates work.
- •Don’t build in fear of OpenAI updates; focus on controllables
- •User behavior should drive pivots (design agent → document intelligence)
- •Avoid obsessing over IPO/VC narratives; build patiently
- •Long-term view: tools may shift from efficiency to fulfillment/pleasure
- •Competing with giants is uncertain—optimize for focus and mission, not rivalry
- 44:43 – 56:12
Bootstrapping and company structure: funding Kuse without VCs
In response to learning context engineering, X.K. pivots to what he believes is more valuable: how they structured the company to stay focused. He explains they used profits from a trading business to fund Kuse, avoiding the distraction of fundraising cycles.
- •Founder background includes YC experience and understanding capital markets
- •Built a trading company first; used proceeds to self-fund Kuse
- •Argues fundraising can distort strategy and consume attention
- •Bootstrapping enabled patience and a “meditative” focus on product-building
- 56:12 – 57:18
Closing recap: context engineering as the core lever behind Kuse’s growth
Aakash wraps by reinforcing the episode’s thesis: context, not clever prompts, drives reliable AI outputs and product value. He points viewers to additional resources (newsletter/podcast links) and ends with calls to subscribe and follow.
- •Reiterates: context engineering is key to non-generic, reliable AI results
- •Highlights Kuse’s traction as proof of the approach
- •Directs audience to extended conversation and linked documents/frameworks
- •Final subscription/follow requests and sign-off
