firehose> #llmops
video youtu.be Watch ↗

160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)

TL;DR

An 18-minute walkthrough whose real content is a distribution claim, not a capability one: because (the video asserts) model capability has converged and the gap between best and worst is under 5%, the leverage has moved off the model and onto the packaged agentics you install into it — and the install path is a one-line git clone of a public GitHub repo into ~/.claude/skills/. Three repos carry the argument: tenfoldmarc/llm-council-skill (five opposed advisor personas — contrarian, first-principles, expansionist, outsider, executor — that answer in parallel, then peer-review each other and emit a council-transcript-*.md whose most valuable section is "Blind Spots the Council Caught"), last30days-skill (fans sub-agents across Reddit/X/YouTube/HN/Polymarket in a bounded recency window and returns ranked evidence clusters with verbatim quotes and exact URLs, shown side-by-side against Claude's un-skilled answer, which is generic and source-less), and Garry Tan's gstack (an entire org chart — CEO, staff engineer, design engineer, QA lead, DX tester, CSO — shipped as /-commands). The load-bearing, mostly-unstated mechanic underneath all three: the model was already capable of every one of these behaviours; what the repo supplies is the method, written down, invoked by name. The video's own screen quietly undercuts two of its framings — the vetting heuristic ("lots of stars = worth your time") is applied to a 366-star repo, and the "one prompt" that built a landing page in 3–4 minutes is a ~130-word spec that names Three.js, glassmorphism, scroll animations, and a QA pass.

Concepts introduced

Held, not dropped (touched by the capture; no concept page warranted yet — spin out on demand):

Key claims

Why this builds on the existing graph

Three concept pages already hold this video's premises, and it reaches them from a different road. Execution Commoditization argues from a cheap model tying an expensive one that convergence is a fact about the task; this video argues from a report's headline number that convergence is a fact about the models. Same conclusion — "it doesn't really matter which model you use anymore" — via a weaker premise, so treat it as a second, independent voice on the conclusion, not as new support for the mechanism. Capability Overhang's "bolting a motor onto the steam-era layout" is precisely "most business owners are still using AI like a chatbot"; two sources now agree, and the video's prescription (install a method, don't just prompt harder) is a small, concrete instance of redesigning the building.

The substantive extension is Adversarial Planning Council. That page was written from a source where the council is a pre-build gate returning Go / Reshape / Kill. This capture is an independent, packaged instance — different author, different repo, same shape — and it refines the concept in two ways. First, the council here adjudicates an operating decision ("raise prices or chase acquisition?"), not a build/don't-build; the pattern generalises beyond the plan-stage gate. Second, and more useful: the artifact identifies where the value actually accrues. Five parallel advisors produce five opinions; the peer-review round produces "what every advisor missed" — unquantified unit economics, an oven-capacity ceiling. Neither was any single persona's answer. That argues the cross-review pass, not persona count, is the load-bearing component, which is a testable claim the original page did not make.

The tension worth writing down. Agentic Distribution holds, as a principle, "distribute agentics by reference, not by copy — copying is the mechanism that produces drift." This video's entire install path is a copy: git clone <repo> ~/.claude/skills/<name>, and updates are never mentioned. last30days-skill had shipped a commit five days before capture and sits at version 3.8.1; a cloned copy is stale the moment it lands. The tension is real but scoped: Agentic Distribution is written about agentics you author and edit across many surfaces, and its own best-practice concedes that below a couple of repos you should "just use plugins / install from wherever." This capture occupies exactly that concession — third-party consumption, no local edits, one machine. So: not a contradiction of the principle, but a boundary the principle should state explicitly. The unresolved half is that the video never tells its ~160,000 cloners how to pull upstream changes, and a git clone-installed skill has no update story at all. Recorded in Public Skill Adoption as an open tension, not settled here.

Finally, the video contains an unforced counter-example to its own vetting advice — recommending a 366-star repo one minute after teaching star-count triage — which is the vault's personal dedup position stated by accident: popularity scores attention, not fit. And its climax, a "one prompt" that is visibly a 130-word spec, is Spec-Driven Development and Imagination Constraint operating unnamed. The video does not notice either. Both are worth holding: the source's demonstrations are consistently stronger evidence than the source's framings, and where they disagree, believe the frame.

Illustrated walkthrough

Visual coverage is ok (274 kept frames, largest un-illustrated stretch ~34 s, 36% grid-floor). No blind gap large enough to hide a section, but frames are sampled — absence of a frame is not absence of an on-screen change, and the sampler is weak on text-on-solid-background transitions.

t=00:00 — the premise, delivered to camera. "Stanford has just released their annual report on AI… two findings." First: "all of the large language models are converging… today, that difference is less than 5% … artificial intelligence has been commoditized." Second: "most businesses have not caught up… most business owners are still using AI like a chatbot." No chart is ever shown — these are asserted, not evidenced, and everything after depends on them.

t=02:15 — repo 1. On screen: tenfoldmarc / llm-council-skill, Public, three files — .gitignore, README.md, SKILL.md. The README opens "Stop trusting Claude's first answer" and "Based on Andrej Karpathy's LLM Council … agents with different thinking styles." Two files of markdown and no code: the artifact is pure method.

t=02:47 — the roster. README continues: "…through 5 AI advisors who argue, peer-review each other … adapted to run entirely inside Claude Code using sub-[agents]." Narration names them: contrarian, first-principles thinker, expansionist, outsider, executor. The video is explicit that Karpathy's original polls different LLMs, whereas this variant runs all five personas on Claude — "but it still works quite well." Diversity of stance substituted for diversity of model.

t=03:14 — the vetting heuristic, undercut by its own screen. Narration: "how many stars and how many forks … when it has a lot of stars, it's an indication that this particular project is very popular. That's another way to help you vet good free projects versus the ones that are just not worth your time." The sidebar in that exact frame reads 366 stars, 4 watching, 44 forks. The heuristic is stated, then not applied to the repo it is stated over.

t=03:54 — the gate. "You need to download the desktop version of Claude. They will not work on the web-based version." claude.com/download, then the Code tab.

t=04:19 — the whole install, one line. Typed into the Claude prompt box (not a terminal):

git clone https://github.com/tenfoldmarc/llm-council-skill ~/.claude/skills/llm-council

Model selector reads Opus 4.8, effort High. This single frame is the video's actual thesis: a capability is a directory, and adoption is a copy.

t=04:57 — invocation and the worked question. "council this: Helena's bakery has a monthly cupcake subscription. Subscribers are loyal, but growth has flattened. Should we focus more on getting new subscribers or raising prices…? Should we either launch a new referral program or should we open a second location? Debate it."

t=05:30–06:06 — the output, and the part that matters. A markdown artifact, council-transcript-20260628-134902.md, laid out per advisor. Visible on screen: The Outsider — "the framing itself hides the real issue… 'Acquisition vs retention' is a choice you make AFTER you know the cause"; The Contrarian — "it also challenged the one assumption everyone else made for free: that loyal customers will tolerate a price increase. Loyalty at today's price is not proven willingness to pay." Then a synthesis section, "Why reasonable advisors split," and — the highest-value artifact in the video — "Blind Spots the Council Caught": "The peer review round was unusually unanimous and surfaced what every advisor missed": unit economics never quantified (no margin, CAC, or LTV), and kitchen/oven capacity as a fulfillment ceiling that could make a referral surge a liability. The value came from the cross-review round, not from the five parallel answers.

t=06:45 — repo 2. …ys-skill (last30days-skill), Public: 46.3k stars, 3.8k forks, MIT, 786 commits, last commit 5 days prior. About: "AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web — then synthesizes a grounded summary." Topic tags include deep-research, recency, polymarket, openclaw, clawhub. A live, maintained project — and the one where the star heuristic would actually have fired.

t=08:00 — the loaded-or-not tell. "If you do not see the last-30-day skill when you type — and this turning blue — that means you have not loaded the skill set properly."

t=09:56 — the output shape. File on screen: ~/Documents/Last30Days/zapier-honest-sentiment-praise- complaints-raw-v3.md, headed "Ranked Evidence Clusters." Each cluster carries a numeric score and a source link: human approval layer for AI agents in Zapier workflows (score 41), "another cash grab from Zapier… why I'm moving to Make" (score 39), "anyone generating PDFs from Zapier without it being a nightmare" (score 34). Verbatim complaints, scored and ranked, each pointing at the URL it came from.

t=11:30 — the controlled comparison, and the real argument. The same prompt run without the skill. Claude's un-skilled answer: "The overall vibe is still the default, but the love is more conditional than it used to be," plus aggregate stats from G2 and Capterra and a generic recommendation. The narration's verdict — "very generic… very bland" — is the only empirical claim in the video, and it is the right one: the difference between the two outputs is not intelligence, it is grounding (dated, quoted, linked evidence) versus recall.

t=13:13 — repo 3. gstack, MIT. README quotes Karpathy ("I don't think I've typed like a line of code probably since December" — No Priors, March 2026), then Garry Tan: "Peter Steinberger built OpenClaw — 247K GitHub stars — essentially solo with AI agents… gstack is my answer. In the last 60 days: 3 production services, 40+ shipped features, part-time, while running YC full-time. On logical code change — not raw LOC, which AI inflates — my 2026 run rate is ~810× my 2013 pace (11,417 vs 14 logical lines/day)." A "LOC Controversy" link is offered pre-emptively. The video reports the star count (115,000) and moves on.

t=13:25 — the roster as a table. Each row is command → role → methodology: /devex-review DX Tester ("navigates docs, tries the getting started flow, times TTHW, screenshots errors… the boomerang that shows if your plan matched reality"); /design-shotgun Design Explorer ("generates 4–6 mockup variants… taste memory learns what you like"); /design-html Design Engineer ("the output is shippable, not a demo"); /qa QA Lead ("fix them with atomic co[mmits], generates regression tests for every fix"); /qa-only QA Reporter ("same methodology as /qa but report only"); /pair-agent Multi-Agent Coordinator; /cso Chief Security ("OWASP Top 10 + STRIDE threat model… independent finding verification"). Note /qa vs /qa-only: the same method, split by whether it may write.

t=15:00 — what installing a roster does to your palette. The / menu is now a scrolling list — bookkeeping, brain-builder, browse, canary, careful, carousel-designer, claude-api, code-review, codex, color, compact, consolidate-memory… The hover tooltip on canary reads "Post-deploy canary monitoring. (gstack) (user)" — origin and invocation mode are surfaced per command. Bulk-cloning a roster is visibly a bulk purchase of operator cognitive load (Skill Invocation Trigger).

t=15:42 — the "one prompt," in full. The prompt box shows /plan-ceo-review; above it, the actual Nexus request: "Build me a landing page for a fictional AI startup called Nexus… Single HTML file, Three.js via CDN, no backend. Dark, futuristic, premium aesthetic: near-black background, deep violet and cyan accents, soft glow, glassmorphism cards… Hero with an animated 3D centerpiece, either a slowly rotating wireframe globe or a flowing particle network of connected nodes that reacts subtly to mouse movement. Add scroll-triggered fade-and-rise animations, a sticky glass navbar, a features grid with glowing icons, animated stat counters, a pricing section, and a clean footer… Then run QA and open it in the browser so I can see it." Claude's reply, and a receipt: "2m 13s · 7.0k tokens."

t=16:31 — the artifact. A dark, violet-and-cyan landing page: "Agents that actually get work done / Not chatbots. Autonomous workers that plan, act across your tools, and report back — 24/7," with glassmorphism cards for "200+ integrations" and "Enterprise-grade security." Narration: "this used to be something that you pay a couple thousand dollars for a UX designer… but now it can be done within minutes with one prompt." The frame shows that the "one prompt" specified nearly every element the page displays. The spec did the work; the model rendered it.

Linked from