Building AI Apps Live from Cairo 🔴
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 focuses on methodical preparation for building an application and an AI-agent command center: tracing tasks back to the manifesto and roadmap, auditing them with several language models, and consolidating the reviews into a canonical task file. Questions about timeline inclusion, dependencies, and how best to judge conflicting model reviews remain partly unresolved. Luqman also works through a task-board design that preserves context, run history, acceptance criteria, guardrails, and verification protocols while assembling task-specific agent instructions. Early prompt previews appeared too long and generic and still required refinement and testing; he likewise expected a large automated implementation to need debugging rather than assuming it would succeed. Wider discussion covers live UI previews inside the development environment, parallel work with agent teams, controlling agent scope, validating task integrity before interface polish, and building without a strong programming background or paid third-party tools. The stream also touches on Shopify submission preparation and the need to connect product building with distribution, marketing, and waitlist strategy. Luqman presents these as observations from a public learning process, arguing that logical thinking and a clear workflow can help while leaving both the build and its eventual market results uncertain.
46 moments from the stream
- 0:01:36
From a Project Map to a Unified Command Center
Luqman explains how the idea evolved from a visual map showing what would be built and when, alongside an AI-agent-operated task board, into a command-center concept that brings the roadmap and task execution together.
- 0:02:30
Preparing Tasks for Agent Execution
He describes a workflow he had used to prepare a week’s tasks in advance: each task carried system instructions, context, tests or a verification protocol, and notes from earlier work, after which an agent could be asked to execute a specific task. He notes that newer automation tools later changed this approach.
- 0:04:23
Validating a Roadmap Before Execution
He discusses an AI-generated multiweek estimate and explains that he does not want execution to begin until the project has been decomposed into correct, complete tasks. The review is intended to catch missing or unnecessary work early rather than after implementation is underway; the duration remains a tool-generated estimate, not a confirmed schedule.
- 0:08:04
Tracing Task Origins and Preserving Deferred Context
Luqman begins reviewing how the application derived its tasks from the planning document, returning to earlier agent sessions and prompts to verify their source. He also asks for the deferred task-board discussion to be saved in a notes file for later, but the verification itself is not completed on this page.
- 0:10:39
Tracing Tasks Back to Project Documents
Luqman investigates which document produced the 165 tasks, asking whether they came from the manifesto or the roadmap. The explanation says checklist items under roadmap milestones were converted line by line into structured subtasks, while work on the task board is deferred.
- 0:12:52
Cross-Project References vs. Duplicating Files
The discussion examines how a command center can access files stored in another project on the same machine. The response recommends keeping the roadmap and its checklists where they belong as a single source of truth, because copying them would make the command center self-contained but create two copies that must be updated together.
- 0:14:59
Designing a Reliable Workflow for a Canonical Task List
Luqman explains that he is unsure whether the roadmap-derived tasks are accurate and asks how to validate them against the manifesto, stabilize most of the swim-lane work, and still allow limited later changes. He considers reviews by multiple language models and asks which documents each model should receive and what single file could become the canonical reference.
- 0:18:40
Auditing Tasks Through Multi-Model Consensus
The response proposes giving each model the manifesto and roadmap, asking it to audit the tasks and identify gaps against the manifesto, then comparing the outputs for consensus to produce a canonical tasks file. It also advises doing this validation carefully once and maintaining the list with discipline instead of building an automated multi-model pipeline at this stage.
- 0:19:50
Should the Canonical Tasks File Include the Timeline?
Luqman asks whether the proposed canonical tasks file should also include the timeline so the tasks can be arranged in the swim lanes accordingly. The supplied page ends before an answer or completion of the thought, so timeline inclusion remains an open question.
- 0:25:21
Reviewing a Project Plan with Multiple Language Models
After following this topic, the viewer will understand Luqman’s described workflow: give Gemini, GPT, and Claude the roadmap and manifesto, collect their proposals for a primary task file, then ask one model to consolidate the different views into a single source of truth. He also notes from his experience that cross-referencing models can become confusing on large files because each model takes a different approach, so complete agreement should not be expected.
- 0:30:40
Validate Task Integrity Before Polishing the Task Board
Luqman explains that he wants to verify the underlying swimlanes and tasks before returning to interface polish on the task board. He flags an unexpected task count as something that needs investigation. The passage shows a practical order of priorities but does not establish the cause of the count.
- 0:31:31
How to Compare Multiple Model Reviews Without Revealing Their Identity to the Judge
Luqman discusses the problem of choosing a judge when comparing three reviews produced by different models. He considers submitting the review files to the models for their opinions, then leans toward using an agent to synthesize and assess them. To reduce the judge’s exposure to model identity, he renames the files so the evaluator cannot tell which model wrote each review. As a personal observation from the experiment, he says Claude produced the cleanest review output and created a Markdown file without being explicitly asked. The passage does not reach a final verdict on the best review or model.
- 0:38:31
Previewing a Website Interface Live Inside the Development Environment
Luqman demonstrates a development-environment feature that opens the website inside the editor so interface changes appear live. He also explains that the embedded preview can select an element on the page and pass it directly to a Gemini agent, which he considers useful for website UI development. His comparison with another environment is presented as an experience-based opinion.
- 0:40:07
Building with AI Without a Strong Programming Background
The speaker explains that he is not a developer with a strong programming background, yet has previously built things with AI and believes experimentation does not require deep technical expertise. A tool comparison is mentioned, but the transcribed names are too unclear to publish reliably.
- 0:41:11
Turning Audit Files into a Unified Task Source
The speaker asks an agent to process several audit files and consolidate them into one structure as a single source of truth. The intended outcome is a task file that replaces tasks scattered across the roadmap and swimlane before work moves to the task board and agent hub; this is an intended result, not a confirmed success.
- 0:43:33
Logic and Technical Background When Working with Agents
The speaker argues that logical thinking and the ability to explain problems to an agent are the foundation, while acknowledging that technical background can help someone move better and faster. His estimate that logic accounts for most of what is needed is a personal opinion, not a documented test result.
- 0:44:09
Combining Building and Distribution
The speaker explains that the channel's approach is not limited to building a project; it also includes distribution and creating a marketing funnel rather than focusing only on building or AI.
- 0:44:24
Designing an Agent-Managed Task Board
The speaker describes a task board intended to be controlled by agents once the swimlane is organized, with each task requiring system instructions, a verification protocol, context, and memory notes. He also explains that an agent was comparing and synthesizing audit files to form a roadmap source of truth and reduce the chance of omitted or unsupported tasks, without confirming that the process completed successfully.
- 0:47:09
Troubleshooting a Background Audio Issue
The speaker tries to determine why the background audio is not coming through, performs an audio test, and checks device settings where one item appears disabled. The captions do not establish a final cause or a successful fix.
- 0:53:45
Turning success criteria into a trusted task file
The speaker contrasts his familiarity with consumer e-commerce and the Shopify ecosystem with his relative inexperience building B2B SaaS companies. He then proposes converting agreed success signals into a tasks.md file that serves as the source of truth for the next implementation step.
Got a question about something in this stream?
Ask in the group