KIMI K3 + TTFD?
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 combines work on a Shopify-focused tool for measuring traffic, sales, and visibility from AI platforms—and improving product pages and FAQs—with practical development of an interactive website that explains the product before installation. Reviews covered page inventory and reusable templates, visual hierarchy, effects, footer details, and cross-device behavior, while acknowledging that some pages, simulations, and fixes remained incomplete or unverified. Luqman argues that AI web-design tools need clear references, a design system, and repeated human direction; his impressions of Kimi K3, Claude, Codex, and OpenCode varied in speed, quality, and instruction-following, but did not amount to a conclusive systematic comparison. The group also discussed the cost and privacy of discounted API providers, key and configuration management, and local model requirements. Explanations for unusually low prices, subscription-versus-API differences, and possible caching benefits remained hypotheses requiring verification and clear data-retention policies. Audience questions covered AEO versus SEO, B2B versus B2C marketing, and evaluating business ideas before building. A recurring conclusion was to prepare the conversion funnel and track metrics such as CAC, LTV, churn, and trial-to-paid conversion before scaling spend, while carefully costing incentives and free periods; unfinished plans and experiments were not presented as proven outcomes.
72 moments from the stream
- 1:04:21
Setting Up OpenCode for a Large Task and API-Key Access Concerns
The speaker explored using a coding tool for a large task and preferred OpenCode or Warp over Claude Code because of concern about the tool accessing an API key. He worked through the setup, considered the task's potential usage, and described Kimi as very slow; these were live impressions, not results from a systematic test.
- 1:12:54
Separate configurations for switching between AI models
The segment discusses setting up access to different AI models, including a custom configuration with a base URL and an additional API key while retaining the previous key so separate configuration groups can be switched. One proposed group is named Codex. Some model names are unclear in the captions, and the segment does not conclusively establish a successful test.
- 1:21:22
Testing an AI Coding Model on an Outdated Pricing Page
A practical test is prepared to assess whether an AI coding model can rebuild an outdated pricing page using a close visual reference. The discussion covers directing another session to the target page and then starting the test, with attention to first-pass accuracy and usage, but this page does not show a completed result.
- 1:30:15
Could Caching Explain Lower Usage Costs?
A tentative suggestion is made that caching might help explain lower model-usage costs, followed by a decision to observe what the experiment does. The page provides no verification or final conclusion for this explanation.
- 1:34:02
Using a Reference Page as a Starting Point, Then Adapting It to the Brand
The speaker explained his preference for starting from an existing page and code with close visual matching, then changing elements to fit the brand and save time. As a personal assessment, he contrasted Codex’s reluctance with Claude’s quicker response to this kind of request, before agreeing to review a similar alternative.
- 1:40:34
Using a Reference Frame as the Minimum Design Standard
After rejecting an output as not production-ready, the speaker asked that every subsequent frame be compared with “Frame 01” and withheld unless it surpassed that reference in structure, color, cleanliness, and overall quality.
- 1:41:50
Creating a Handoff Prompt for a New Session
The speaker requested a handoff prompt that would let another session resume exactly where the current work stopped, with a specified confirmation word at startup.
- 1:44:01
Animation Ideas for Static Side Elements
The speaker requested a brainstorm on animating elements along the left and right sides of the page because they were static and occupying space without enough visual activity.
- 1:45:55
Reported Repeated Stalling in Kimi K3
The speaker noticed that work had stopped, clarified that he had manually stopped one session, and said Kimi K3 had stalled for him on previous occasions. This was a personal observation, not a systematic test.
- 1:59:08
Retrying a Webpage-Building Prompt with Another Model
After an earlier model did not follow the request as intended, the speaker switched to another model and reused roughly the same steps and prompt. He asked it to inspect the repository and prior examples to recreate a reference page with its effects, but this page does not show a confirmed final result.
- 2:06:53
Rebuilding Web Interfaces and Claude’s Differing Subscription/API Behavior
The speaker explains that he had previously used AI tools to visually reconstruct reference sections and pages, then changed their content and structure for his own design. He observed that Claude through an API/CLI sometimes refused similar requests that had worked in the subscription experience and questioned the reason for the difference. He also noted that visual replication may work while functionality and UX still need attention, and acknowledged copyright considerations.
- 2:14:29
Testing a Prompt to Recreate a Page
The host tested a prompt intended to have an AI model recreate a page as-is, with its content to be replaced later as needed. He treated the result as an experiment rather than assuming it would succeed.
- 2:20:46
Reviewing Generated Frames and Effects
The host reviewed visual outputs and rejected most of them, saying that only one frame was acceptable so far and that one effect felt unsuitable.
- 2:22:11
How Memory Affects Local Model Use
Answering an audience question, the host explained that memory capacity materially affects how comfortably models can run locally. He also relayed another person's suggestion that combining multiple machines could enable stronger models, but the caption ends before the Kimi K3 example is completed.
- 2:25:37
When to Build Custom Skills vs. Install Existing Ones
The speaker says his most useful daily skills are custom-built around a particular repository and project, while established third-party skills make sense when they are well known or costly to recreate in time.
- 2:26:25
Extremely Low Model-Service Pricing: Questions About Economics and Data
The speaker discusses unusually low API and subscription prices and makes a partly unclear cost comparison involving options or models including Kimi K3 and Codex. He speculates—without claiming certainty—about a possible connection between the low pricing and data collection, mentions partial- and full-capacity options, and says his trial was still early with no clear difference noticed yet.
- 2:31:07
Recreating a Visual Effect from a Reference Frame
The speaker asks to place a frame on a dark background for visual inspection, then describes a seemingly simple task: reproduce the complete effect of a reference page in his own page, with manual refinement expected afterward.
- 2:39:32
What Needs to Be Completed Before Launch
The speaker lists the website funnel, creative assets, advertising funnel, email sequences, and email marketing as elements to complete before launch.
- 2:41:01
Using a Template for a Large Set of Pages
After pausing to inspect pricing, the speaker concludes that the number and size of the remaining pages make a reusable template the most practical way to accelerate the work with Codex.
- 2:43:56
Inventorying and Categorizing Website Pages Before Building Templates
A website-redesign workflow begins by reviewing the current branches and worktrees, listing navigation pages in a clear table, and categorizing the large page set before defining templates. The proposed approach is to create several reusable designs so later page work mainly requires inserting content, while the homepage, scan/landing page, and pricing page would receive standalone designs.
Got a question about something in this stream?
Ask in the group