Renderhaus @Renderhaus_

Started with a plain text dump of book titles and page counts. Asked for a local reading tracker: list of books, a progress bar per book, and a total pages left number that updates when you log what you finished. 1.6T wrote the page, the progress math, and a small form to add pages read. Server came up. First book showed the right bar. Second book with a higher page count dragged the total left number the wrong way. I didn’t paste the error. It refreshed, logged pages on the second book, watched the total stay high, then fixed the sum and checked the same book again. Two rounds after the first paint. The page stayed mostly intact; it only touched the running total. My daily model did the math correctly on the first try but rewrote the list component and lost the “mark finished” button. This one kept the button and only missed the total until it looked at the screen. The 397B multimodal build got the running page. It opened localhost, picked a book, typed a page number, hit update, and watched the bar and the total. It caught that one book’s bar had jumped past 100% while the total still counted the full length. One-line fix after that. Same session I pointed it at a browser word-search for about a minute. It scanned the grid and clicked a word. Same motion as the tracker: read the screen, act, check the result. Fine for QA. Not something I’d use for games. 35B on a single 24GB card, Q5, roughly 22 GB and mid-30s tokens per second. I gave it a test that the total pages left drops by the number you log. It fixed the test and the function in one round. When I asked it to build the whole tracker it wrote the math, then stalled on the list layout. That’s the pattern with the small one: solid on the isolated function, weak once the UI has more than a couple of moving parts. Cheap enough that I just keep it for the tests. Time on the full tracker was about 19 minutes with the 1.6T plus a few extra minutes while the multimodal model clicked through updates. Daily model was a bit slower because of the rewrite. The 35B never got a usable page. The useful split is still the same: small local model for the pure math tests, bigger text model to turn a list into something that runs, multimodal model to actually update a book and tell me when the total is lying. If it can’t see the number on the screen, the tracker isn’t finished.

九月 @tuttihq 的数据，截到 9 月 30 日： ▸ 累计注册达人 14,269 个，九月新增 3,528。四月的时候总共才 159 ▸ 真正发了帖的 1,431 位，一共 61,601 条，分在 46 场活动里 ▸ 九月达人奖励 $81,480，1,545 个人拿到。本月单人最高 $2,653，累计拿得最多的那位已经 $11,197 ▸ 有效曝光 860 万 ▸ 一个月上了 7 次 𝕏 热榜 ▸ 达人的粉丝里美国占 25%，抽样到的粉丝分布在 189 个国家和地区 十月接着冲。。。

@Stellakjbk this beats screenshot + paste into chat every single time

@Stellakjbk adding voice wake-up next would make this feel complete

国区 iPhone 没有 Apple Intelligence / Siri AI？那就自己接一个。 我用 iOS 快捷指令 + OpenRouter，把 Ling 3.0 Flash VL 接进了 iPhone。 现在直接拍一张图，就可以交给视觉模型识别、理解并返回结果。视频里随手拍了一页 NVIDIA 的新闻稿，几秒后就能得到中文摘要、关键信息、风险点和建议。 相当于给国区 iPhone 手搓了一个自己的「视觉 Siri」👀 后面准备继续加：截图识别、上下文追问、语音唤起。 OpenRouter现在可以免费使用Ling-3.0-flash-VL： openrouter.ai/inclusionai/li… #AI #iPhone #AppleIntelligence #Ling3

This thread is saving me from another launch with just static images because I've been in the exact spot of having a solid product ready but no good way to show it dynamically – prepping branded screenshots and letting Pexo handle the motion, timing, and multi-format output while keeping everything on-brand is the practical workflow indie creators like me need to stop skipping video entirely and actually get better engagement on our tool releases.

沐阳 @yyyole

a month ago

从0-1讲透产品Launch Video，入门到精通（附提示词） x.com/i/article/2102…

465 311 469 203K 328

This thread nailed why I'm getting more into agentic AI tools for content creation – you kept feeding Pexo feedback on shot density, rhythm and information reveal timing until the 30-second piece had proper narrative weight instead of just flashy visuals. I've been experimenting with similar loops for my tool demos and it saves so much time compared to traditional editing while keeping your personal vision intact, especially useful for indie creators who wear multiple hats.

九尾狐-FoxFairy🦊 @Stellakjbk

a month ago

x.com/i/article/2102…

370 257 293 120K 75

Your iterative process with Pexo on the asteroid escape sequence really shows how treating AI video like an actual director conversation changes everything – starting from a rough reference and refining the fake-out safety moment into real tension made the whole SpaceX piece feel alive. As someone building my own explainer videos for productivity tools, this back-and-forth on pacing and story beats is exactly the workflow I need to stop getting generic outputs and start creating with more control.

x.com/i/article/2102…

This ASTRA concept film workflow with Pexo feels like the practical upgrade I’ve been looking for in my creator stack – start with the shape transformation on a simple black background, then use Mark to Fix in the same chat to hold the glow and accents. I’ve been testing similar loops for productivity tool demos and it cuts the usual restart frustration so I can stay focused on story instead of fighting the output.

Vox @Voxyz_ai

a month ago

𝗠𝗮𝗱𝗲 𝘁𝗵𝗶𝘀 𝗶𝗻𝘁𝗿𝗼 𝗮𝗻𝗶𝗺𝗮𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗔𝘀𝘁𝗿𝗮. Made with @Pexoai_offical . It’s an unofficial concept film. The video shows the finished animation, how I made it, and one revision. After the first version, I used 𝗠𝗮𝗿𝗸 𝘁𝗼 𝗙𝗶𝘅 to mark the area around

359 252 294 124K 87

What really clicked watching this is how you turned an abstract geometric idea into a polished concept film by using Mark to Fix in-chat to lock the title glow instead of regenerating everything – as someone deep in building systems for creator productivity and juggling multiple AI experiments daily, this precise control loop addresses the biggest pain point I face with video tools: keeping vision intact across revisions without losing momentum. The shared prompt structure from shapes to name reveal on that simple black background is super practical for indie makers like me who need fast yet professional assets for launches, and it's got me excited to test a version of this workflow soon on my upcoming feature updates to get better engagement without the usual production drag.

Vox @Voxyz_ai

a month ago

𝗠𝗮𝗱𝗲 𝘁𝗵𝗶𝘀 𝗶𝗻𝘁𝗿𝗼 𝗮𝗻𝗶𝗺𝗮𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗔𝘀𝘁𝗿𝗮. Made with @Pexoai_offical . It’s an unofficial concept film. The video shows the finished animation, how I made it, and one revision. After the first version, I used 𝗠𝗮𝗿𝗸 𝘁𝗼 𝗙𝗶𝘅 to mark the area around

359 252 294 124K 87

I've spent way too much time trying to organize my research, bookmarks and half-finished ideas into something coherent for threads or tool demos, so watching Pexo generate this full brand-matched animation from their site alone really highlights how these AI tools are solving that persistent context problem most of us face daily; it keeps the original voice intact while making the concept click visually, which has me planning to feed my own site in soon to create better promo assets without the usual editing grind.

Pexo @Pexoai_offical

a month ago

Hey @supermemory , we made this for you. Your website was the only input. Pexo picked up your brand style and turned what you’re building into a visual story.Scattered context, brought together by a memory layer for AI. Does this feel like Supermemory to you? #MadeWithPexo

375 271 358 193K 94

Been deep in testing different AI tools to organize my scattered research and project ideas lately but always end up with disconnected pieces that don't flow well together, so seeing Pexo turn their website into this clean visual breakdown of the connective memory concept feels like the practical solution I've been missing for creator workflows; the brand voice match is spot on too and could save so much time when I need quick explainers for my own tools without hiring help or learning video editing.

Pexo @Pexoai_offical

a month ago

Hey @supermemory , we made this for you. Your website was the only input. Pexo picked up your brand style and turned what you’re building into a visual story.Scattered context, brought together by a memory layer for AI. Does this feel like Supermemory to you? #MadeWithPexo

375 271 358 193K 94

This video nails the exact struggle I hit every week when trying to explain my own productivity setups to people – Pexo pulling a full animated story from just the Supermemory site without losing the core memory layer idea is impressive and makes me want to test it on my messy note system next.

Hey @supermemory , we made this for you. Your website was the only input. Pexo picked up your brand style and turned what you’re building into a visual story.Scattered context, brought together by a memory layer for AI. Does this feel like Supermemory to you? #MadeWithPexo

@arya_at1 @AnthropicAI This thread shows the mature way to use these tools: clear boundaries, specific constraints, and staying inside the existing system.

The flaky test only failed after deploy. Local was green. CI was a coin flip. That is the job I gave Claude Code from @AnthropicAI. Same repo. Terminal in the project, not a fresh chat tab. The bug: a Next.js route handler wrote the session cookie after a redirect had already started. Preview looked fine. Production dropped the user on the second hop. Playwright caught it once in five runs and nobody trusted the suite. Brief I typed: do not rewrite the auth stack do not add Redis do not “just use middleware for everything” set the cookie before the redirect fix the test so a dropped cookie fails every time then run it It read the handler, the test, and the CI log. Moved the cookie write above the redirect. Tightened the Playwright assertion so a missing Set-Cookie is a hard fail, not a retry. First run still flaked on timing. It waited for the response header instead of guessing at a sleep. Second run was boring. Boring is the point. What I like about Claude Code is the loop stays in the repo. It will over-explain the auth model if you let it. I cut that paragraph. The patch was two files. That is the whole review. I would not use it to invent login. I used it because this was a header-order bug wearing a “we should rebuild auth” costume. Same rule as the other agents I keep: current behavior, expected behavior, files it may touch, things it may not invent, then run. @AnthropicAI — the useful version of this tool is not the essay. It is the cookie that lands before the 302, and a test that fails when it does not. If your session needs a new platform, you do not have a redirect bug. You have a meeting.

keyboard_arrow_left Previous keyboard_arrow_right Next

@arya_at1 @githubCopilot @github The fact it fixed the temp dir permission issue in the second run shows it was really operating inside the actual repo.

The issue was labeled good-first-task. It was not. A rate limiter on a public route used a global Map, so two instances doubled the quota and one deploy wiped the counts. That is the job I gave @githubCopilot. Not autocomplete. The agent on the issue, in the same @github repo. Brief I left on the ticket: keep the existing middleware do not add Redis do not invent an API gateway store hits per instance without lying across deploys add a test that fails if two processes share a counter open the PR when the test is green It read the issue, the middleware, and the flaky test. Replaced the Map with a file-backed counter that dies with the process on purpose — honest local limit, no fake cluster story. Wrote the two-process case. First CI run failed on path permissions. It fixed the temp dir, pushed again, stopped. I deleted a comment about “eventual consistency for rate limits.” We did not have a distributed system. We had a Map. What I keep Copilot for is GitHub-shaped work. Issue already exists. Review sits on the PR. Checks run where the code already lives. I would not use it to design the product. I used it because this was a process-local bug wearing a “we need infra” costume. Same brief I use everywhere: current behavior, expected behavior, files it may touch, things it may not invent, then run. @githubCopilot @github — useful when the agent can be assigned the ticket and the proof is the check, not the essay. If your rate limit needs a new platform, you do not have a counter bug. You have a meeting.