AI COMMERCE INDEX SCORING 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 documents a build-in-public effort around an engine for measuring how visibly Shopify merchants’ products appear in AI answers, alongside a Command Center for classifying agent tasks by size and dependency, executing them, and reviewing them through task- and milestone-level audits. Practical work covered code-informed prompt preparation, correcting illogical dependencies, monitoring tasks and logs, and improving the README and installation options. Although some workflows were described as becoming more autonomous, defect review, the frontend, and several implementation decisions remained unfinished. Luqman also compared Codex, Claude, and other tools from personal experience, consistently qualifying observations about speed, quality, resource use, and session variability as non-benchmark impressions rather than settled conclusions. Audience questions and wider discussions covered market validation and marketing before Shopify approval, organic outreach versus paid acquisition and waitlists, landing-page structure, animated backgrounds and calls to action, and GEO/AEO compared with traditional SEO. Other useful threads included choosing Telegram or Discord for an Arabic community, learning programming, AI, and DevOps through practical projects, data residency, hosting and DNS, plus ideas for automating marketing and short-form clip extraction. A recurring conclusion was that building a product does not automatically solve distribution; multi-platform expansion, the pre-launch strategy, Hermes adoption, and parts of the website and domain work remained open or under investigation.
123 moments from the stream
- 0:02:07
Classifying Tasks by Execution Mode
A command-center workflow classifies tasks as fully human, agent-executed, or human-agent pair work, with the observation that paired work can require more time than delegating a task to an agent alone.
- 0:03:31
Scoring Store-Product Visibility in AI Tools
The speaker describes building a feature that scores how visibly a store or merchant’s products appear in answers from tools such as ChatGPT, Perplexity, Gemini, and Claude.
- 0:06:24
A Practical Impression of AI Coding Models
Answering a question about Chinese models, the speaker says that after trying several alternatives, his personal hands-on experience did not reveal anything stronger than Claude and Codex. He frames this as a practical impression rather than a benchmark-based conclusion.
- 0:08:48
Automating a Milestone Sweep by Task Complexity
The speaker describes a command-center sweep that works through a selected milestone whose tasks are labeled small, medium, or large according to complexity.
- 0:09:34
Building an Arabic AI and Vibe-Coding Community
The speaker says his primary aim goes beyond channel growth toward a community whose members exchange knowledge and help one another, and expresses the view that Arabic interest in AI and vibe coding is worth developing and has a future.
- 0:11:31
A Command Center for Autonomous Task Execution
An explanation of a command center that accepts a high-level command, inspects the work queue, and selects small tasks that can be executed. The speaker also said he planned to document the commands in a README and publish the application on GitHub so users could customize it to their workflows.
- 0:13:33
Ordering Task Dependencies and Unblocking Work Automatically
A practical discussion of small tasks being blocked by prerequisite medium tasks, with a proposal to number tasks and assign logical labels based on dependencies. According to the explanation, the system is intended to wait for prerequisites, then approve the work, pass it to an auditor, and automatically unblock the next task.
- 0:18:48
Coordinating Parallel Work Across Development Streams
The speaker discussed running work in parallel across the command-center repository, the main AI Commerce Index project, and its website, using dependency numbering to identify tasks that can proceed together. He noted generally that token and rate limits can constrain how much agent work runs in parallel.
- 0:22:46
Preparing Implementation Prompts Through Research and Code Exploration
The speaker described a workflow that takes medium-complexity tasks from a development milestone, uses research and codebase exploration, and then writes a separate prompt for each task. He said these detailed prompts make implementation easier because they are grounded in prior project analysis.
- 0:26:20
AI Visibility Scoring Engine for Merchants
The speaker explained that the project is building a scoring engine for Shopify merchants to estimate how visible their products are in systems such as ChatGPT. He compared the idea to search optimization within AI environments and noted that the work is divided into tasks of varying complexity.
- 0:26:55
Different Workflows for Tasks by Complexity
The speaker explained that small tasks can run together, while large tasks require step-by-step handling. He said he believed small tasks might be completed automatically after entering review. In his description of the system, medium and large tasks enter a review path where an auditor agent examines the code and tests before the task is considered complete.
- 0:29:16
Parallel Work Without Disrupting the Primary Task
The speaker argued that progress can be faster when independent, non-overlapping parts of a project are developed in parallel. He also noted that resource limits or difficulty maintaining focus can constrain this approach, so he preferred to protect the primary task from disruption.
- 0:32:39
The Roadmap as the Foundation of an Agent Command Center
The speaker argues that a precise, carefully developed roadmap is essential for an agent command center’s swimlanes and task board. He suggests giving agents a comprehensive project manifesto, asking them to analyze it alongside the repository, and producing a system-specific phased roadmap while checking for potentially omitted tasks.
- 0:37:41
Automating Short-Clip Extraction from Streams
The speaker explains that he sends a full stream to StreamLadder so AI can select short clips automatically, reducing the manual time required to prepare shorts and reels.
- 0:39:11
Building the Product Does Not Solve Marketing
The speaker advises that a project built with AI tools may be struggling because of marketing rather than product construction. AI can assist with marketing but does not deliver it automatically, and specific advice requires enough context about the project.
- 0:43:22
How Codex and Claude Differ in Agent Workflows
The speaker inspects how Codex and Claude prepared prompts for project milestones and observes different approaches to reasoning, work organization, and token usage. He compares their prompt and file structures while explicitly remaining uncertain about which methodology is better, making this an exploratory observation rather than a final verdict.
- 0:51:56
Establishing a Single Prompt Source for the Builder
The review found that the builder reads its prompt from the tracker record while separate human-facing Markdown files contain different wording. After considering which source and methodology were preferable, the proposed direction was to wire the builder to execute directly from the Markdown prompt files, with an implementation plan if the change proved substantial.
- 0:54:27
How Interface and Tooling Shape the Coding-Agent Experience
Responding to an audience question about perceived differences between Codex and Claude, the speaker suggested that much of the experience may come from the interface, workflow, and availability of tools, plugins, and skills rather than model capability alone. He qualified this as potentially subjective and noted that the tools may embody different working philosophies.
- 1:02:08
Why the Same Coding-Agent Task Can Produce Different Results
The speaker observes that running the same task or prompt in separate sessions can produce different completion times, responses, and detected issues. He considers this variability common across coding agents, while acknowledging that he does not know its precise cause.
- 1:04:29
A Qualified Comparison of Coding Agents by Results and Value
The speaker recounts one poor experience with another model but qualifies the judgment because his testing was limited and the outcome may have been incidental. In his view at the time of recording, Codex offered practical coding value, while he generally preferred Claude but considered it more resource-intensive for coding work.
Got a question about something in this stream?
Ask in the group