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 centers on developing a website and an AI-assisted coding workflow with tools including Codex, Claude, and Antigravity. Practical work covers interface hierarchy, CTA wording, project cards, live preview, and selecting page elements inside the development environment. Luqman also discusses a command center under construction for project maps, tasks, agent teams, and execution prompts generated from project context; inspection suggests the dynamic prompts remain too long and generic and need smarter, role-specific context profiles. Another major thread explores connecting the website to YouTube live status and separate live/offline thumbnails through options such as the YouTube Data API, Cloudflare Workers, and KV, but the integration was deferred and not shown as complete. The stream also covers automating content repurposing, organizing YouTube playlists, and separating recommended or affiliate-linked tools from the technical stack. In audience discussions, Luqman argues for learning by building while continuing college and stresses starting with a real problem, validating demand and unit economics, and planning distribution before implementation. A Telegram community and an initially manual paid-membership verification process are considered, while questions about automated GitHub pushes and permissions remain unanswered. Much of the work is exploratory or unfinished, and the session includes recovery from a power outage and a website crash whose cause was not established.
40 moments from the stream
- 1:14:11
Generating Execution Prompts from Project Context
A workflow starts from a short task and has an agent inspect the project context, manifesto, and relevant files to generate a more detailed execution prompt before running the task. The aim is to reduce the need to hand-write every agent instruction.
- 1:15:54
A Plan to Make the Command Center Reusable
The speaker discussed a future plan to polish the command center and make it available to others, allowing it to read each project’s context and manifesto and automatically create a suitable workflow. These capabilities were presented as work in progress, not a completed release.
- 1:17:26
Different Thumbnails for Live and Offline Stream States
After confirming that the play button opens the stream, the discussion turned to adding a static thumbnail for the link, with one image while live and another while offline. Creating a manual design was mentioned as an option, but implementation was not shown on this page.
- 1:20:53
Discord or Telegram for an AI-Building Community?
The speaker says Discord appears less widespread in the Arab world than abroad and suggests Telegram as a possible home for a small community where vibe coders help one another.
- 1:25:00
From Vibe Coding to a Sellable Product
The speaker defines vibe coding as using AI to build websites or applications without a programming background. His business advice is to begin with a real problem, plan customer acquisition and distribution early, and build a waitlist and interest before launch rather than postponing sales until the product is finished. He notes that finding customers remains a common challenge and that the right plan depends on the product.
- 1:28:20
Refining a Website Section with Codex
The speaker begins using Codex to improve a website’s “What I’m Building” section. He asks for the main project to be more prominent and for clicks to open a restrained explanatory popup instead of expanding the item inline. The page captures the request and the start of the work, but does not establish that the change was completed or successful.
- 1:31:57
Learning by Building Instead of Relying on Comparison Videos
The speaker argues that short tutorials and conflicting AI-tool comparisons can confuse beginners. He recommends starting a real build and learning through practice, while using occasional videos as supporting material.
- 1:33:48
Context-Window Use in AI Coding Tasks
The speaker compares an estimate of context-window use for a refactoring task with an earlier one-shot website-generation attempt, noting that such tasks can consume a substantial portion of the context window.
- 1:36:19
Live Streaming as a Tool for Focus and Reflection
The speaker explains that talking through work aloud during a stream helps him focus and hear his own reasoning, while recordings provide a way to review progress from earlier weeks or months. He also tentatively considers inviting guests into future streams.
- 1:37:16
Separating Recommended Tools from the Tech Stack
The speaker distinguishes a recommended-tools section from the project’s technical stack, asking that various services be discussed in the appropriate section rather than mixed with the technologies used to build the project.
- 1:38:37
Prompting a Coding Agent to Refactor Interface Hierarchy
The speaker drafts a refactoring request that places a command center at the top of the hierarchy and visually links it to related projects, then notes that the monetization model should be planned before the section is completed. The page does not establish that the implementation finished or succeeded.
- 1:40:56
Market the Tool’s Value, Not Personal Earnings
A suggestion to frame marketing around the value users receive and encourage them to try the tools, brainstorming several approaches instead of emphasizing the seller’s personal earnings.
- 1:45:23
Question About Automating GitHub Pushes with a Custom Command
A question was raised about creating a custom command that automatically pushes changes to GitHub and handles settings or permission constraints. This page does not contain a confirmed answer or implementation.
- 1:45:51
Learn to Build AI Products While Continuing College
The speaker advises a student to continue college while using spare time to build real things rather than mainly watching tutorials. He argues that AI tools can reduce the need for an extensive programming background and suggests consulting an assistant such as GPT or Claude to relate interests, strengths, and student status to a suitable direction; this is the speaker’s opinion, not a career guarantee.
- 1:51:37
Improving an AI-Generated Interface with a Stronger Prompt
Luqman reviews an interface still in progress and says its appearance is disappointing despite using a strong model setting. He concludes that it needs a better prompt, then asks to remove the guides section and explain how a paid Telegram offering differs from regular Telegram and can be monetized technically.
- 1:54:20
Verifying Members for a Paid Telegram Offering
The discussion considers setting up a paid Telegram offering, with manual verification presented as reasonable at an early stage until membership grows, followed by a request for exact setup guidance. The subsequent captions are fragmented, so no specific automation method is attributed.
- 1:59:32
Reviewing and Preparing a Stream Page
Luqman reviews the above-the-fold portion of the stream page and plans two Canva thumbnails—one for the live stream and another for the static version—while leaving the broadcast window and some page sections for later refinement.
- 2:01:11
Diagnosing and Improving an Agent Task Board’s Prompt Generator
A walkthrough of an agent task-board concept where the agent writes a task prompt from the project’s current state, history, context files, manifest, and workflow lanes. Inspection suggests the assembled prompt is dynamic but overly long and generic, with broad manifest sections and sibling tasks included without enough task-type differentiation. The speaker discusses needing a stronger system prompt and smarter context selection, then leaves the work running while an incomplete proposal mentions agent profiles and better sibling-task control.
- 2:04:54
Push Permissions and a Slash Command
A practical question asks whether allowing automatic pushes and removing the denying setting would let a slash command work. The page does not provide a confirmed answer or test result.
- 2:11:13
Role-Specific Context and Prompt Profiles
A design for profiles that control prompt and context layers by agent type: an explorer receives a minimal brief and skips the broad manifesto. It also proposes generating a structured multiline prompt when no custom prompt exists, showing a budget indicator in the UI, and leaving the IPC contract unchanged. Manual steps were suggested to preview and confirm a task prompt, but this page does not clearly establish that the test was completed.
Got a question about something in this stream?
Ask in the group