RIP. NEW MISSION. CC V2.0
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 redesigning a largely autonomous “Command Center” that routes coding tasks by size and dependency, delegates research, exploration, and implementation to agents, and introduces task- and milestone-level audits before human review. Luqman demonstrates unfinished practical work across Codex, Claude, and terminal workflows, weighing output quality against token consumption and usage limits while encountering real friction involving Supabase, Docker, MCP, and stalled tasks; the workflow is therefore not presented as a completed success. The discussion broadens to unit testing, lightweight feature specifications, competitor research before building, and the need to market products rather than stopping after implementation. He also outlines a Shopify tool concept for measuring and improving product visibility in AI answers through GEO, which he described at the time of the 2026 recording as an emerging field and forecast to grow, particularly toward late 2026, while leaving market viability, implementation complexity, and the GEO–AEO distinction unresolved. Drawing on experience that began with dropshipping in 2017, Luqman compares dropshipping with DTC and white labeling, e-commerce with app economics, and B2B with experimental B2C directions; he also offers opinions on SaaS margins and recurring revenue, reasons for choosing the Shopify ecosystem, and app marketing through defined audiences, acquisition channels, and a mix of organic and paid approaches, without presenting these views as guarantees or settled outcomes. Secondary discussions cover automated short clips and the risk of lost context, recorded streams as a work archive and community-learning channel, programmatic marketing videos with Remotion, thumbnail experiments, and copyright constraints. Audience questions add useful discussions about learning software development through practical AI-assisted projects, verifying the security and provenance of agent skills, estimating time, managing dependencies, model pricing and compute efficiency, and scheduling work around session limits. The qualified conclusion is that automation may reduce manual coordination but does not remove the need for oversight, debugging, and validation; Luqman argues that strong planning and context matter more than minimizing token use alone.
155 moments from the stream
- 0:00:20
Choosing Work Hours Around Productivity
The speaker discusses starting streams and work in the morning despite stronger nighttime engagement because he finds mornings more productive and wants to change his late-night routine.
- 0:03:54
A Lesson in Competitor Research Before Building
After discovering a free alternative he considered stronger than his earlier idea, the speaker reflects on researching the market and competitors before building and treating setbacks as signals to reconsider direction.
- 0:06:23
Improving E-commerce Product Visibility in AI Engines
The speaker outlines a proposed e-commerce tool that would score visibility in AI engines, detect and address product-data issues, and track outcomes attributed to that channel. He presents it as part of the emerging GEO field, not as proven results.
- 0:09:01
Cleaning Up a Project and Rebuilding the Workflow After a Pivot
The speaker explains that a product-direction change led him to reorganize his command center, remove unused code, and establish a new workflow suited to the revised scope.
- 0:10:40
Choosing a Practical Workflow Across Codex, Claude, and the Terminal
The speaker gives a personal comparison of Codex, Claude, and terminal-based work. He considers Claude stronger for his needs overall, but expects to use Codex more because of its available usage capacity, and praises Codex’s visual tools for selecting webpage elements and requesting direct changes. For now, however, he chooses their CLI interfaces inside a familiar terminal environment that supports multiple sessions and terminals.
- 0:14:47
Question About an Auto-Research Tool and Token Cost
Answering an audience question about an auto-research tool whose name is unclear in the captions, the speaker says he has heard of it but has not used it. His tentative understanding is that it is a plugin-like tool that automates research, while its token consumption remains an open question; he asks users with firsthand experience for feedback.
- 0:16:16
Redesigning Command Center Around Time Awareness and Agent Efficiency
The speaker explains that week-based dates were removed from the workflow because, in his experience, language-model task-duration estimates were unreliable, while broader time awareness remains in the system. He also says the previous version was slow and token-intensive because it deployed multiple subagents per task. The redesign uses an MCP server connected to CLI tools and a set of commands to drive the workflow; he then begins introducing lessons about debugging and review before this source segment ends.
- 0:20:02
Why Manual Approval of Agent-Generated Code Falls Short
The speaker explains that when agents deliver large amounts of backend and UX-logic code, manual review becomes impractical, especially for a non-specialist, so the workflow needs an additional review layer.
- 0:22:35
Integrating Automated Research into an Agent Workflow
The speaker discusses making research mandatory for medium tasks and suggests that an automated-research approach attributed to Karpathy could fit the command center’s closed loop. He considers using Perplexity through MCP or an API, while noting that the referenced GitHub repository did not specify this and token usage remains a constraint.
- 0:24:17
Task and Milestone Audit Agents Before Final Debugging
The speaker describes a workflow that can commit, push to GitHub, and trigger deployment, then adds one agent to audit each task against a defined checklist and another to review an entire milestone, including validation and security. The aim is to catch issues before final human debugging, while acknowledging that debugging will still be necessary and audits consume tokens.
- 0:26:54
Commands for Preparing Milestone Tasks in Parallel
The speaker explains that tasks are classified by size to drive the workflow, then demonstrates custom commands such as next for inspecting current work and prepare for launching research and exploration agents to prepare a selected task class within a milestone. The segment ends with an unresolved practical question: whether the command will include blocked tasks or only open ones.
- 0:30:02
Executing Task Batches with Coding Agents
The speaker explains a workflow that prepares a batch of medium-capacity tasks, delegates exploration and research to agents, and executes available work while dependent tasks remain blocked. Tasks then progress through review and auditing, with completion unlocking downstream work. He believes this can reduce manual interaction and the need for multiple terminals, while requiring close token-usage monitoring; the reported results are his experience, not an independent test.
- 0:32:13
Balancing Coding-Agent Quality, Usage Limits, and Cost
The speaker compares Codex, Claude, and other model options in terms of work quality, usage limits, and cost. He expects to use Codex more because of constraints he encountered while still preferring aspects of Claude's quality, and suggests evaluating other models for learning or specialized work such as UI tasks. He also offers the opinion that higher prices may reflect compute costs and service demand, while remaining uncertain whether an earlier usage change was a bug or an intentional change.
- 0:39:28
Cost Expectations for DeepSeek V4
The speaker gives his personal expectation that DeepSeek V4 could help lower prices by delivering stronger performance at a lower compute cost. The discussion is brief, and the source ends before the explanation is completed.
- 0:43:36
Orchestrating Coding Tasks Across Coordinator, Builder, and Auditor Agents
A practical explanation of a workflow in which a coordinating session launches agents for exploration, research, and implementation, followed by an auditor reviewing the work before completion. It also discusses narrowing the preparation milestone and monitoring token use while selectively cleaning up an old store and retaining necessary parts.
- 0:46:03
Could Compute Efficiency Lower AI Prices?
An audience discussion considers whether memory techniques and reuse of stored facts could reduce the computation needed for simple questions. The speaker presents competing views on future prices and leaves the outcome uncertain, tying it to compute utilization, supply and demand, and continued growth in AI adoption.
- 0:50:43
A Cautious Discussion of DeepSeek Model-Cost Claims
The speaker discussed circulating claims that DeepSeek models were inexpensive to develop, while framing them as uncertain personal recollections and acknowledging limited expertise in AI backend technology. He broadly believed the company used an approach different from other models in the market.
- 0:54:35
Automatically Advancing Tasks Through Dependencies and Status
The speaker explained a workflow where completing one task can unblock another small task and let the system continue automatically. He described it as highly autonomous while retaining human oversight for blockers, dependencies, and final project review, and noted that usage limits could affect throughput.
- 0:57:45
Why the Backend May Shape the Frontend
The speaker considered starting with the UI and frontend, then rejected that order for this work because backend decisions could shape the interface. He said reverse engineering the backend from the frontend was possible but could introduce problems, and suggested that a more autonomous workflow might reduce his need for multiple agents except during debugging or final stages.
- 1:00:11
Can Work Be Scheduled Around Usage-Limit Resets?
Responding to a question about continuing work automatically while the user is away when usage limits reset, the speaker suggested investigating hooks, loops, and dispatch as possible event-driven mechanisms. He clearly stated that he did not know of a dedicated solution and did not establish that the idea worked.
Got a question about something in this stream?
Ask in the group