Fixing Broken Code
Want to back the stream? The «داعم» membership on YouTube is EGP 60 a month: badges, emoji, and I answer your questions first on stream. Become a supporter
The gist
This portion centers on developing a Shopify app and preparing it for submission, with a brand “manifesto” presented as an optional profile that can guide content personalization across the product. The discussion covers questionnaire design, terminology and approval actions, manifesto visibility, and locked research states, while pushing back on feature creep in favor of finishing backend wiring, debugging, mobile review, and compliance work. The build and final testing were not completed: upload and draft actions were tested without a documented final result, while unexpected localization behavior and a batch-update bug were identified. The speaker stresses human visual QA after each milestone because automation alone had allowed bugs to accumulate. This portion also explores coding-agent workflows, including session handoffs, multi-agent orchestration, unresolved sequential-versus-parallel tradeoffs, environment contention and wait time, and why large refactors should not be assumed to finish autonomously. Audience questions prompt cautious opinions on B2B versus B2C pricing, acquisition and conversion economics, two-sided marketplace growth, advertising-platform feasibility, and researching profitability before building. Additional discussion touches on marketing automation, learning through practice, agent-tool choices, and trying different AI models, while keeping economic claims, model-capability inferences, and interpretations of AI-policy reports explicitly uncertain or opinion-based.
65 moments from the stream
- 0:54:08
A Two-Phase Shopify Compliance Review
The speaker plans to research the latest Shopify compliance requirements and then sweep the codebase for risks and potential issues. The process is divided into a knowledge and research phase followed by an execution phase based on the findings.
- 0:56:14
Simplifying Manifesto Review and Approval Actions
The discussion covers prompting users gently to verify their answers without diminishing the manifesto’s value, while allowing them to return to earlier questions. It also explains the tradeoff between a short initial questionnaire that reduces user effort and the depth needed for a strong manifesto, and proposes replacing an ambiguous CTA with a clearer approval action.
- 1:00:35
Manifesto Approval Flow and App-Wide Impact
A discussion of returning users to the intelligence page after manifesto approval and showing the manifesto’s downstream influence on articles, products, optimization, research, and the wider application instead of relying on shallow descriptive pills.
- 1:11:01
Limits of Inferring Model Capabilities
The speaker argues that available signals may support estimates of some competing-model capabilities, but are unlikely to reveal its core mechanism directly, leaving substantial room for guesswork.
- 1:12:13
Improving the Manifesto Generation Introduction
A specification update calls for cleaning up the page shown before manifesto generation and creating a stronger mockup whose introduction clearly communicates the purpose and encourages users to generate it.
- 1:15:19
Plans to Try AI Models Across Varied Tasks
A discussion about plans to try additional AI models across varied projects and tasks, including website design and debugging, with the aim of developing more practical experience using them. Some model names and details are unclear in the captions and are therefore left unidentified.
- 1:19:24
Designing States for a Locked Research Feature
A review of how to present a locked research feature as an in-page upgrade, with mockups requested for its unavailable state and its state after research is generated. Private plan and usage specifics are omitted while preserving the general state-design concept.
- 1:23:10
Simplifying Compliance Review and Measuring Readiness
The speaker asks for a complex analysis to be summarized in plain language and scored out of 100, using that score to judge readiness before submitting an application and aiming for a very high level of compliance.
- 1:24:50
Reviewing Landing-Page Layout and Product Messaging
The review covers a desktop layout with two vertical cards, the risk that “Category Scout” could be misunderstood as referring to a merchant’s product categories, and the need to reframe the wording. The speaker also considers highlighting the research capability—either naming Perplexity or describing the capability alone—and asks whether the broken attachment interface is already implemented or still being brainstormed.
- 1:29:54
Stopping Feature Creep to Focus on Debugging
The speaker believes the application is sufficiently ready and decides to stop adding features and focus on debugging after continual additions caused fatigue.
- 1:37:22
Turning Deep Research into Focused Research Hunts
The speaker discusses reframing a deep-research feature as separate “hunts,” each focused on a lane such as the market, niche, or product. The aim is to lower the cost per task, support more tasks, and use their findings to guide agents inside the application. He also weighs fixed subscription pricing against pay-as-you-go billing, noting that usage-based billing may add implementation complexity and is not currently essential.
- 1:39:46
A Cautious Reading of a Government Report’s AI Scope
The speaker says that, in his reading, the report did not explicitly call for shutting down a particular system; instead, he believed its argument might also extend to systems from OpenAI. He presents this as an uncertain personal interpretation, then expresses hope that Anthropic succeeds because its tools are important to his work.
- 1:45:18
Visualizing the Workflow Before a Coding Session
While preparing for a coding session, the speaker asks the assistant to lead the discussion, propose an end-to-end workflow, and present it visually before implementation continues. He also notes that expanding the feature set has created additional workflow-design work.
- 1:56:51
Defining executable work and the manifesto PDF specification
The speaker proposes resuming document compilation and executing ready items, deferring work that depends on backend wiring, and specifying the manifesto PDF’s structure and approximate page count.
- 1:58:25
Connecting revenue forecasts to acquisition cost and conversion
Answering a question about future profits, the speaker says pre-market forecasts are uncertain and depend on marketing investment. He explains that acquiring an install does not guarantee a subscription because users receive a seven-day trial, so acquisition costs and conversion must be considered together. The conversion-rate explanation is incomplete at the page boundary.
- 2:06:15
Value and Pricing in B2B and B2C Markets
Based on his prior experience as a Shopify merchant, the speaker argues that business customers tend to seek value and may accept higher prices, while consumer products generally need lower pricing. This is his market assessment, not a demonstrated test result.
- 2:08:09
Parallelism Versus Sequential Execution
The speaker asks what parallel execution offers over sequential work: whether it is faster or more efficient, and how those benefits compare with increased token consumption. This page does not contain the answer.
- 2:11:05
Dynamic Multi-Agent Orchestration
A brief remark says that an unclearly named coding mode spawns orchestrators and multiple agents by default, and suggests avoiding a fixed cap on agent count. The model or tool identity is uncertain from the captions.
- 2:11:24
Including Visual Mockups in Agent Instructions
A practical question about embedding visual mockups in the working document because they provide important direction to agents, especially when code does not render as expected.
- 2:12:13
Calculating Cost per Install for Ad Campaigns
An explanation of why cost per install should be calculated when evaluating ads, noting that Meta advertising covers Facebook and Instagram while the speaker qualifies that TikTok ads are outside his strong expertise.
Got a question about something in this stream?
Ask in the group