Inside the Core Engine
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
The stream centers on developing the Burnish Shopify app, especially refining its multi-step onboarding so it feels native and consistent with Shopify, communicates merchant value clearly, and improves copy, cards, navigation, and pricing using visual references. The work remained unfinished: content and design revisions, backend wiring, deep-link behavior, and onboarding’s dependency on database state still required attention. In parallel, Luqman explores a safer Git workflow for multiple coding agents—cleaning the repository, isolating tasks with branches and worktrees, detecting conflicts before merge, validating development servers and URLs, and maintaining a living diagnostic log of difficult problems and fixes. Audience questions and broader discussions cover SEO and GEO, AI-driven product discovery, using funnel signals to assess demand, and practical advice for learning while building a SaaS product. The stream also contains personal, qualified impressions of Codex, Claude, Gemini, GPT, and coding benchmarks, stressing that outcomes depend partly on how tools are used and that informal comparisons do not support firm conclusions.
109 moments from the stream
- 4:42:26
When a Screenshot Is Part of the Specification
The discussion explains that treating a screenshot as decoration beside the prompt caused the written wording to take precedence when the two conflicted, while the visual evidence was ignored. The correction was that literal replication had been requested clearly and that the implementation needed a more precise reading and matching of the design elements.
- 4:44:09
Preparing an Application for Visual QA
Before visually reviewing the interface, the development servers needed to be checked and an accessible application URL provided, with attention focused on the onboarding sequence under development.
- 4:47:05
An Onboarding Flow That Blocks the Rest of the App
The speaker explains that the onboarding flow prevents work on other areas and blocks normal access to the application, while bypassing it may require resetting onboarding state in the database. This highlights how a mandatory flow can impede development and testing.
- 4:48:30
Diagnosing Background Differences in a UI Replication
While comparing the implementation with a reference screenshot, a concrete mismatch was identified: a missing white background behind the cards that visually connected them. The discussion stresses inspecting visual relationships and overall structure, not only words or isolated elements.
- 4:49:54
How Frustration Degrades Prompt Quality
The speaker observes that becoming frustrated while working with an AI tool makes the prompt worse, reducing the tool's ability to help and making the interaction more difficult.
- 4:51:08
Workstreams Before Submitting an App to Shopify
A progress update outlines several workstreams before submitting the app to Shopify: completing onboarding, finishing the interface design, and addressing the internal engineering or functionality expected to take longer. A fourth workstream is mentioned, but the caption ends before it is explained.
- 4:52:11
Refining Onboarding UI and Moving Work Between Coding Tools
After an earlier change worked, the next request was to match the cards’ corner radius to the primary CTA. A handoff prompt was requested to explain the active branch and the goal of completing all four onboarding steps, while Codex was tried because another tool did not fit the branch- and worktree-linked preview setup.
- 4:53:24
Creating Prerecorded Content with Remotion and Codex
The discussion described using Remotion with Codex to create content for prerecorded lectures. Remotion was presented as powerful for this work but requiring experience and having a learning curve, even if it is not inherently difficult.
- 4:59:35
Why Do the United States and China Lead in AI?
Answering an audience question, the speaker said he believed other countries also develop AI, but that the United States and China are far ahead because of deeper investment. He framed competitive model development as a major economic and national investment decision, and cautiously mentioned efforts in Europe and the UAE without identifying a specific model.
- 5:01:52
Context Compaction and Browser Integration Limits in Codex
The speaker explains that when the context window fills during a task, the conversation is compacted and a recap of prior context is passed forward. This generally works but may weaken continuity. He also contrasts the agent’s limited access to an authenticated browser session with another tool, while noting that running Codex inside the development environment makes screenshots easy to provide.
- 5:06:27
When Onboarding Depends on the Database
The speaker explains that work inside the application is temporarily blocked because its onboarding flow depends on database state required for activation. He therefore wants to finish and polish that flow before moving to other work.
- 5:07:00
A Simplified, Qualified View of How Language Models Work
Responding to an audience question, the speaker offers a non-expert account: language models perform numerical computation and predict the next word from prior context. He repeatedly says he does not know the training details or whether methods have changed, so his comments about the effects of coding training are presented as attributed impressions rather than established facts.
- 5:20:18
Workflow Bottlenecks from Sequential Agent Sessions
The speaker explains that GitHub optimization followed by onboarding work required waiting for each task to finish before moving to other sections, making the workflow constrained and somewhat tedious.
- 5:21:40
Evaluating Agent Runtime Against the Actual Task Scope
After questioning a long wait for what appeared to be a small request, the speaker discovered that the agent had implemented the full onboarding sequence rather than only one section. He therefore reassessed the runtime as relatively reasonable, while noting an available faster mode and the need for follow-up revisions.
- 5:23:41
Reviewing an Onboarding Sequence and Identifying Revisions
The review showed that the multi-step onboarding sequence had been completed, but some content and design still needed changes, with particular concern about a later step before the work could be considered finished.
- 5:31:46
Opinions on Using GPT and Gemini for Different Tasks
The speaker said GPT works well for his coding workflow, while he mainly uses Gemini for deep research and brainstorming. He qualified that his coding experience with Gemini is limited and that his negative CLI assessment was based on what he had heard, not a documented personal test.
- 5:34:17
Refining Onboarding Timing and Messaging
An onboarding review proposes emphasizing that AI-search store optimization starts immediately, restoring a short timing message, and centering the sentence beneath the illustration. The discussion notes that onboarding may take roughly one to two minutes, with the result visible when the user reaches the application.
- 5:38:48
Using Real Screenshots Instead of Approximate Graphics
The speaker considers improving the output by giving the tool appropriate access to screenshots. He tentatively believes the current image may have been recreated from scratch as an SVG based on the expected appearance rather than taken from a real screenshot.
- 5:40:33
Improving a product step and capturing the live store URL
The speaker reviewed a second product step, considered its design broadly sound with minor changes, and said its content should address a genuine need. After examining a competitor’s approach, he questioned whether capturing the live store URL served an internal application function. He discussed reliably prepopulating the live store URL and questioned how this should behave for a password-protected development store without a live site.
- 5:45:16
Avoiding distraction across AI tools
The speaker discussed enthusiasm for new models while preferring to focus on one or two rather than chase every AI tool. He connected FOMO to the pace of competition around Shopify apps in the same GEO-related niche, then offered the qualified opinion that Google could become a strong coding competitor if it concentrated its resources there.
Got a question about something in this stream?
Ask in the group