Working on Merchant Facing UI
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 follows preparations for a Shopify app submission, prioritizing completion of the move from Polaris React to Polaris Web Components and visual and functional QA of the merchant interface before returning to website, report, and funnel polish. Practical work also covered GA4, Google Tag Manager, and PostHog configuration; verifying that changes reached production through GitHub and Cloudflare; reviewing Supabase table organization; and troubleshooting MCP access—without treating visible identifiers alone as proof of end-to-end tracking. Luqman drew lessons from accumulated Git branches and worktrees and from coding-agent completion claims that failed manual inspection, arguing for regular integration and direct verification. He also discussed automated build-and-audit agent loops, his personal impressions of Codex and Claude, and audience questions about deployment, model choice, project languages, client acquisition, programmatic SEO, Kanban boards, and Hermes. Open questions remained around some Cloudflare and MCP failures, the best database organization, the value of particular agent tools, and the final submission date.
104 moments from the stream
- 2:35:55
Verifying Task Completion and Web Component Adoption
The agent was asked to explain the task changes and verify completed items, after which the review was said to confirm that the Shopify application had fully adopted web components.
- 2:37:20
Handing Off a Milestone to Another Agent Session
Requirements were set for a handoff prompt to another agent session: first complete and close the milestone, then open the merchant-admin interface and jointly perform visual QA across the surface.
- 2:38:45
Clarifying Verification of Production HTML
The discussion asks what should be verified in deployed HTML, corrects that the latest deployment was to production rather than preview, and requests a step-by-step walkthrough of the confirmation process.
- 2:42:44
Diagnosing Missing Analytics Configuration
An attempt to determine why the GA4 script, Google Tag Manager, measurement ID, and a PostHog project key were not visible, with parallel investigation proposed and escalation advised for Cloudflare or DNS blockers.
- 2:43:58
Practical Comparison of Coding Agents for Tasks and Deployment
The speaker gives a personal assessment of a large difference between two AI coding tools when completing and deploying work, while questioning whether his own instructions may have affected the outcome.
- 2:49:50
Merchant Admin UI Review and Visual Verification
A review of ongoing Merchant Admin Service work in which several tasks require manual UI review and visual verification, followed by opening the Shopify interface to continue design work with AI-assisted oversight.
- 2:51:29
Choosing Reasoning Effort by Task
The speaker compares Codex and Claude for UI work and explains that he varies Claude’s reasoning-effort setting with task difficulty while tending to use a high setting for Codex; these are personal judgments, not a documented benchmark here.
- 2:53:50
Choosing a Reasoning Level and Using GPT Through OpenCode
The speaker suggests that high reasoning may be enough for clear problems, while the extra-high setting can take longer and has sometimes failed in his experience. He also explains, based on his understanding, that OpenCode accesses GPT through its API, so GPT does not become a different model inside OpenCode.
- 2:55:07
Verifying a Coding Agent’s Work Instead of Trusting Its Confirmation
The speaker describes an experience where an agent reported that an analytics and deployment task was complete, but manual inspection showed that the expected changes were absent and deployment had not occurred. He frames this as a reliability and tooling problem, noting that a Supabase plugin worked for him when MCP access did not.
- 2:59:12
Handing Work Off When a Coding Agent Stalls
After the agent struggles to connect to project services and complete a merge, the speaker asks it either to finish the work or provide a handoff prompt for another agent. He emphasizes that the comparison used the same project and access to the same skills and MCP servers, while leaving the reason for the behavioral difference unresolved.
- 3:02:20
Adding GitHub Repository Variables as a Quick Workaround
The speaker locates the GitHub repository variables or secrets settings, adds the variables, merges the change, and asks the agent to verify and continue. He qualifies that repository-level configuration was a quick workaround and not necessarily the best approach.
- 3:07:12
Comparing Two Models in the Same Project Environment
In an audience discussion about organizing Markdown files, the speaker says two models used the same repository, files, skills, MCP servers, and plugins but performed differently in his experience. He raises, without resolving it, the possibility that the project’s organization may align more closely with Claude’s working style than GPT’s.
- 3:09:30
Why Markdown Files Do Not Replace MCP Servers and Plugins
The speaker explains that MCP servers, plugins, and Markdown files are not direct substitutes: files provide context and instructions, and a skill may be built around a file, while MCP and plugins provide a bridge to external tools or services. He therefore questions the suggestion to disable them all and rely only on Markdown files, while acknowledging that he may be missing part of the argument.
- 3:14:58
When Automatically Triggered Agent Skills Become Distracting
The speaker discusses how an agent may automatically invoke a skill or plugin even when the request does not need it, adding workflow steps and token usage. He says this behavior is usually visible, while other failures may instead come from different tool or website access across agent environments; this remained his hypothesis rather than a confirmed diagnosis.
- 3:17:34
Investigating a Cloudflare Deployment Failure
The speaker describes a case where one assistant quickly deployed the same repository to Cloudflare but later struggled with or declined a similar step. The discussion considers deploying from a local Worker file rather than GitHub, differences in runtime access, and the need for environment variables to exist in the main project so they appear in Cloudflare, while leaving the final cause unresolved on this page.
- 3:23:09
Verifying analytics configuration after deployment through GitHub and Cloudflare
The discussion explains that a direct deployment did not expose the expected Google Tag Manager, GA4, and PostHog configuration because required changes had not gone through GitHub. The changes were merged and deployed, then the identifiers were checked in the page source before the task was marked complete; this alone did not establish that end-to-end event tracking was working.
- 3:28:27
Improving AI coding-tool integration with the workflow
The speaker reflects that another tool could have completed the deployment more directly, but argues that the better long-term approach is to configure Codex more effectively for the existing tools and workflow so it can handle deployment and merging more reliably.
- 3:32:14
Moving on to review the application UI inside Shopify
After resolving the measurement-configuration issue, the speaker reprioritizes toward reviewing the application interface inside Shopify and performing visual QA on the dashboard as the next important task. This page does not contain the results of that review.
- 3:33:20
Setting Up Shopify Preview and Logs for UI QA
The speaker explains that visual UI QA requires running a Shopify app development server inside the coding environment, opening the app preview, and keeping the logs readable while using another browser.
- 3:35:00
Speculation About Model Modes and the Execution Harness
In a brief audience discussion, the speaker speculates that effort or speed modes offered by coding tools such as Cursor with Claude and Codex may relate to the execution harness under discussion, without claiming a firm conclusion.
Got a question about something in this stream?
Ask in the group