{"apiVersion":"v1","generatedAt":"2026-10-25T04:22:21Z","pulse":{"slug":"ai-influencers","name":"AI Influencers Pulse","description":"What leading AI voices are saying and sharing on X.","sections":["Tricks & Techniques","Worth Knowing","Analysis & Insights"],"itemCount":20,"latestEditionItemCount":3,"lastEditionAt":"2026-10-25T04:00:00Z","feedUrl":"https://presto-ferry-g7bz.here.now/feeds/ai-influencers.xml","itemsUrl":"https://presto-ferry-g7bz.here.now/api/v1/pulses/ai-influencers.json"},"count":20,"items":[{"id":"4d0172af27d8d2b2","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Stretch your Claude limits with /model opusplan","url":"https://x.com/jayzhoupro/status/2114192036840780168","synopsis":"Responding to the day’s big complaint that unlimited Opus 5.5 use is over (pcshipp’s post drew about 1,600 likes), jayzhoupro points to a built-in fix: in Claude Code, /model opusplan runs Opus while planning and Sonnet while executing, and switching models doesn’t wipe the conversation. I confirmed the opusplan alias in Anthropic’s Claude Code model-configuration docs, but couldn’t find the quoted “fastest way to drain a daily limit” line on that page, so treat that wording as unverified. To try it, type /model opusplan at the start of your next session and save full Opus for genuinely hard, wide-reaching changes.","also":["https://docs.claude.com/en/docs/claude-code/model-config","https://x.com/pcshipp/status/2114041082480935133"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":1},{"id":"a155c27fda03125b","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Why personal agents feel magical to newcomers but not to daily coding-agent users","url":"https://x.com/dotey/status/2114199763121844347","synopsis":"Asked why bosses and even veteran programmers are getting hooked on personal agents like Grok Bot and Muse, dotey (writing in Chinese) compares it to the earlier “little crayfish” agent craze (likely a nod to OpenClaw), only with a lower barrier since nothing needs installing or configuring. His read is that people who haven’t used a smart model paired with an agent get a sci-fi jolt plus real emotional comfort from a “personal assistant,” while people who use Claude Code or Codex every day barely react because they’ve already seen what good looks like. It’s a useful lens for judging agent hype: ask whether the excitement comes from first exposure or from capability that’s actually new.","also":[],"section":"Analysis & Insights","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":2},{"id":"c6cff260f042b984","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Hazel, an open-source talking Halloween candy dispenser built on Claude","url":"https://x.com/RoundtableSpace/status/2114193984402657487","synopsis":"A 3D-printed witch uses a sonar sensor to spot trick-or-treaters, has Claude write her replies and ElevenLabs speak them, then drops candy, with code that limits it to one piece per child, three per group, plus a cooldown. The original project is religa’s GitHub repo (Arduino, parametric build123d models, sized for a Bambu A1 mini), created Oct 23 and still tiny at two stars; per the post, the README notes visitors’ words go to both APIs, suggests a sign at the door, and a soak test is still open. It’s a fun weekend build if you have a printer and a week before Halloween.","also":["https://github.com/religa/halloween-candy-dispenser"],"section":"Worth Knowing","publishedAt":"2026-10-25T04:00:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T04:00:00Z","rank":3},{"id":"7f582b205fea288d","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"An open-source “AI designer” skill file behind a 2,000-like landing-page reel","url":"https://x.com/elayadesigns/status/2114049200749175174","synopsis":"elayadesigns says every landing page in the attached reel was designed over the past two months with Higgsfield, Claude and Cursor driven by a custom design skill, and published that skill openly; the post drew about 2,100 likes and 3,050 bookmarks, a very high save rate. somi_ai read the file and points out it runs 403 lines with “never” appearing 24 times (never Inter, never italics, no gradient backgrounds, no “Learn more” buttons), which is the real lesson: good taste is mostly a list of refusals. To try it, load the skill into Claude Code or Cursor for your next landing page, or borrow its banned-patterns list for your own design rules. The original is from earlier today (about 10:40 AM PT), is mainly a video, and resurfaced on the list this hour.","also":["https://github.com/elayadesign/ai-design-skills","https://x.com/somi_ai/status/2114178670994878523"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":1},{"id":"f1722bb3adc68197","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Give Claude a CLAUDE.md map of your Obsidian vault","url":"https://x.com/cyrilXBT/status/2114182675108106320","synopsis":"cyrilXBT’s tip is to put a CLAUDE.md file in the root of your Obsidian vault that tells Claude where new notes go, which tags are real, and which notes it must never touch, because Claude reads that file every session instead of guessing differently each time. It’s quick to try: write ten lines covering folder rules, your tag list and off-limits notes, then compare Claude’s next filing job against the last one. The linked X Article from earlier today is pitched as a “full course with templates,” so treat that part as promotion; the core tip stands on its own.","also":["https://x.com/cyrilXBT/status/2114057378001747980"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":2},{"id":"f15d8db6f9176ee4","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Fancier coding-agent harnesses cost up to 3x more without solving more tasks","url":"https://x.com/ds_wen_/status/2114186564217806931","synopsis":"ds_wen_ shared VentureBeat’s write-up of a Shandong University study that held models fixed and ran nearly 500 SWE-bench Verified tasks through Claude Code, OpenCode and the roughly 100-line mini-SWE-agent: Claude Code and mini-SWE-agent landed within two points of each other, while cost per task varied up to threefold, largely because the first call alone carried 16,581 tokens of system prompt and tool schemas in Claude Code versus 829 in mini-SWE-agent. The practical lessons are to run your own baseline twice before trusting a 2–3 point harness claim, trim tool schemas, and tune thinking budgets per model-and-harness pair; the article also cites a matching UC Berkeley “HarnessTax” result. The post itself got almost no engagement, but I read the article and the evidence is solid.","also":["https://venturebeat.com/orchestration/a-coding-agents-harness-can-triple-its-cost-without-making-it-more-accurate-study-finds"],"section":"Analysis & Insights","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":3},{"id":"640ea6e715388ce5","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"A hands-on pushback on the Opus 5.5 hype versus Codex","url":"https://x.com/haider1/status/2114176368120373301","synopsis":"haider1 says Opus 5.5 often commits to an approach, gets stuck, skips researching better alternatives, and brute-forces the same weak design into bugs and optimization problems, while Codex on 6.1 high or ultra researches options much better. It’s one developer’s experience rather than a benchmark, but it drew 10 replies of real discussion and names a testable difference: give both the same refactor and watch which one revisits its plan.","also":[],"section":"Analysis & Insights","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":4},{"id":"df2c37b063ff5ba6","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"The RTX 4090 “2x tensor unlock” is real on paper, smaller in practice","url":"https://x.com/HowToPrompt__/status/2114178962859721157","synopsis":"HowToPrompt__ is spreading gpu-unlock-ada, an open-source Linux kernel module that flips an SM speed-select setting before NVIDIA’s driver loads and roughly doubles FP16 and BF16 tensor throughput with FP32 accumulation on the RTX 4090 (165.6 to 331.6 TF/s). I checked the repo: that figure is register-only instruction throughput on one card, and the one independent matched run found real cuBLAS matrix-multiply gains of 12–34% for FP16 and 16–20% for BF16, with stability testing still pending and the 4070 Ti SUPER path marked experimental. It builds on the earlier Ampere unlock this pulse covered and runs GPU firmware with a modified payload, so only try it on a machine you can afford to break.","also":["https://github.com/JustAResearcher/gpu-unlock-ada"],"section":"Worth Knowing","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":5},{"id":"e50495a322fd27ed","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"A creator says Kimi’s K3.1 promo brief told influencers which narrative to push","url":"https://x.com/elshayib_/status/2113734602032312731","synopsis":"elshayib_ posted what they say is Moonshot AI’s creator campaign document for Kimi K3.1 Pro and said they declined to take part (about 1,600 likes and 145 replies). choblin29 summarized the brief’s “Narratives You Can Consider” as including lines like “China just made Opus 5.5 free to use for everyone,” and AISafetyMemes amplified it this hour, adding that the brief tells influencers not to disclose the posts as paid. I couldn’t independently verify the document or the non-disclosure detail, and AISafetyMemes’s claim of state control is speculation; the useful takeaway is to weigh influencer cost comparisons for Kimi K3.1 with care.","also":["https://x.com/choblin29/status/2113939080203645198","https://x.com/AISafetyMemes/status/2114182278553498089"],"section":"Worth Knowing","publishedAt":"2026-10-25T03:00:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T03:00:00Z","rank":6},{"id":"d0ac4c609dcaa709","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"SaaStr's agent-led growth playbook, with real numbers","url":"https://x.com/jasonlk/status/2114167023924064560","synopsis":"jasonlk's new write-up shows how SaaStr runs go-to-market with 3 people and 21 agents: one queue (they call it the Prospector) for every inbound, outbound reply, renewal and support ticket, where the agent checks first-party data before Exa, Sumble and CoreSignal enrichment, and a real inbound got a meeting booked within 10 minutes. Two plays you can copy today: ask an agent for the top three or four competitors of every customer you've ever had and feed that list to outbound, and run a cold, logged-out AEO audit (they had Muse ask ChatGPT, Gemini and others 50 category questions, and SaaStr came up in 46). The numbers are SaaStr's own and some named tools are vendors they use, but the steps are specific and first-hand.","also":["https://www.saastr.com/the-agent-led-growth-playbook-a-10-minute-inbound-booking-a-46-of-50-aeo-score-and-why-first-party-data-beats-every-signal-tool-the-agents-017/"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":1},{"id":"5d01331b10519a8e","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Bot fleets: one primary that routes, nothing goes out without approval","url":"https://x.com/jerrod_lew/status/2114166049176498201","synopsis":"jerrod_lew joined a chain of people posting their Grok Bot setups: 12 bots covering email, projects, finance, X tracking, AI news, images, and travel and health admin, with one primary bot routing everything and a rule that nothing is sent without his OK. He's quoting pressmanc, who started two days ago, and the chain traces back to a post describing about 44 single-job bots with a router, a \"Fleet Nurse\" that audits routines for cost and staleness, and handoffs through a shared folder. Engagement is small (a few likes and bookmarks each) and the setups are mostly screenshots, but the pattern is easy to copy: give each bot one job and an explicit list of what it shouldn't do, and keep a human approval step on anything outbound.","also":["https://x.com/pressmanc/status/2114039851217715414","https://x.com/i/status/2114026843527659689"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":2},{"id":"bcd4cc3049b2cda1","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Test your moat by swapping to a cheaper model","url":"https://x.com/jayzhoupro/status/2114161225533952192","synopsis":"Riffing on jack's four-word post \"there is no moat\" (about 7,400 likes), jayzhoupro proposes a concrete test: point your product at a cheaper model and re-run your evals. Whatever still works (the workflow, the data that piles up per customer, the reason people renew) is yours; whatever collapses was the model doing the work, and the next release hands that to whoever undercuts you. It's a tiny post with almost no reach, but it turns a vague debate into an afternoon experiment if you already have evals.","also":["https://x.com/jack/status/2114121702515114234"],"section":"Analysis & Insights","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":3},{"id":"6bb163678e0c71f2","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"LLM Dojo, a free course and interview bank for AI engineering","url":"https://x.com/somi_ai/status/2114161848618824095","synopsis":"somi_ai shares LLM Dojo by adxtyahq, a free hands-on course that runs from foundations and inference through prompting, evals, retrieval, agents and production, and suggests starting with evals because without them \"you can't tell if your agent got better or just got lucky.\" I checked the site: lessons are free without an account (signing in adds spaced-repetition review), and it includes about 492 interview questions sorted by level; I couldn't confirm the post's 15-hour figure. The post is mainly a video and has almost no engagement yet.","also":["https://llmdojo.dev"],"section":"Worth Knowing","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":4},{"id":"906747c57b3e27d0","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Claude Code plus Godot builds an open-world RPG from grabbed assets (unverified)","url":"https://x.com/jurlycat/status/2114170928007659874","synopsis":"jurlycat highlights a developer who connected Claude Code to the Godot game engine and gave it browser access to pull premade assets from Sketchfab, Mixamo and Itch, producing GANG WIZARD, a deliberately silly medieval open-world RPG. The post is mainly a video and doesn't name or link the original developer, so I couldn't trace or verify it, and there's no setup detail beyond the tool combo. It's here because the combo (coding agent plus game engine plus asset libraries) is worth trying; check each asset's license before shipping anything built this way.","also":[],"section":"Worth Knowing","publishedAt":"2026-10-25T02:00:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T02:00:00Z","rank":5},{"id":"3647153be9d4df10","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Magazine-style diagrams from one prompt in Claude Code or Codex","url":"https://x.com/kirillk_web3/status/2114157494650257431","synopsis":"kirillk_web3 points to diagram-design, a free open-source skill for Claude Code and Codex that draws editorial-looking flows, architecture maps, timelines and org charts instead of default gray Mermaid boxes; the post says it can pull brand colors and fonts from your website, check label contrast, redraw old draw.io and Mermaid files, and export HTML, SVG and PNG. I confirmed the GitHub repo exists and describes itself as 45 diagram types in self-contained HTML and SVG (the post says 44), but I couldn't confirm the post's 53K-star claim, and the post is mainly a video. To try it, install the skill from the repo and ask your agent for one diagram of something you'd actually put in a deck.","also":["https://github.com/cathrynlavery/diagram-design"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":1},{"id":"7992e2b574ce469e","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Better AI video camera moves with start and end frames","url":"https://x.com/jerrod_lew/status/2114145065795010793","synopsis":"jerrod_lew's tip is to take your start frame, re-render it from a new camera angle, and attach that as the end frame, which gives the video model far more context about the move you want. He shows his workflow in Magnific, but the idea works in any video tool that accepts start and end frames. It's a small post (7 likes) and the demo is visual, but it's a concrete technique you can test in a few minutes.","also":[],"section":"Tricks & Techniques","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":2},{"id":"9dc4e19bb87339fc","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"A ready-made project manager prompt for Grok Bot","url":"https://x.com/Michael_Fenech_/status/2113207232514994264","synopsis":"Michael_Fenech_ resurfaced his Oct 22 setup prompt this hour: it tells Grok Bot to find active projects across connected tools, record each one's goal, deadline, next milestone and owner, check progress daily, draft human follow-ups for approval, and send one daily briefing split into MOVED, STALLED and NEEDS MY DECISION, plus a weekly review. The original has 77 bookmarks, and it explicitly tells the bot not to invent progress or change deadlines without approval. To try it, paste the prompt from the original post into Grok Bot and check the routines it proposes before approving them.","also":["https://x.com/Michael_Fenech_/status/2114156990003818740"],"section":"Tricks & Techniques","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":3},{"id":"abed7bbdc1239b8d","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Bitcoin Core ends long embargoes for bugs an LLM can find","url":"https://x.com/MartinSzerment/status/2114150579605778932","synopsis":"MartinSzerment explains that on Oct 22 Bitcoin Core created an \"easily discoverable\" category for flaws findable with minimal effort, such as a simple LLM prompt, and those now get fixes on every maintained version right away with the advisory published alongside, instead of embargoes that could last about a year. His point is that secrecy only delays the fix for honest node operators when attackers have the same prompt, so the burden moves to people running nodes to update fast. I checked this against Bitcoin Magazine's and Unchained's reporting on Antoine Poinsot's merged policy change and it matches; critical bugs are still handled case by case.","also":["https://bitcoinmagazine.com/news/bitcoin-core-ends-long-embargoes"],"section":"Analysis & Insights","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":4},{"id":"656c31345b1e2313","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"Kindle modding repos now include setup instructions written for AI agents","url":"https://x.com/MartinSzerment/status/2114144178850467959","synopsis":"MartinSzerment noticed that some Kindle modding repos now ship a README section aimed at AI agents: plug the e-reader in over USB, give your agent the repo link, and ask it to do the setup. He ties it to adriandlam_'s Oct 18 post about doing exactly that, which is circulating again. It's a small but telling sign of projects writing docs for agents as well as people; I didn't check which repos do this, and modding can void a warranty or brick a device.","also":[],"section":"Worth Knowing","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":5},{"id":"ea731f32b9f6ddeb","pulse":"ai-influencers","pulseName":"AI Influencers Pulse","title":"A browser skateboarding game built in about ten days with Claude Code","url":"https://x.com/RoundtableSpace/status/2114148685521076694","synopsis":"RoundtableSpace shared Cruising Bones, a free downhill skateboarding game in the browser whose maker says he started Oct 14 and posted his third update about ten days later, using Claude Code and Claude for the code, Magnific and Tripo for art, and BabylonJS as the engine. It's labeled an alpha with multiplayer still coming, and the timeline and tool list are the builder's own account. It's a useful real-world look at the stack a solo developer can use for a 3D web game (18 likes, about 14,700 views).","also":["https://cruisingbones.com","https://www.reddit.com/r/ClaudeAI/comments/1xe7kwg/built_a_downhill_skateboarding_game_with_claude/"],"section":"Worth Knowing","publishedAt":"2026-10-25T01:00:00Z","firstSeenAt":null,"editionAt":"2026-10-25T01:00:00Z","rank":6}]}