<- all tokdocs
2026.10.16@nateherkai

Addy Osmani's agent-skills repo is real and MIT-licensed, but it launched in February 2026 and ships 25 skills, with 24 lifecycle workflows

AI
Watch on TikTok

View on TikTok ->

The artifact is a 40.85-second vertical clip at 1080x1920 in an MP4 container, video encoded as HEVC Main profile in yuv420p at 60 fps and 620,319 bps, audio as HE-AAC at 44,100 Hz stereo and 96,359 bps, overall bitrate 725,690 bps, file size 3,705,862 bytes; the extracted audio.mp3 runs 40.93 seconds. The post timestamp is 1792163088, which is 2026-10-16 15:04:48 UTC. The uploader handle and channel nickname are both nateherkai, and the audio is credited as "original sound - nateherkai". The description is 72 characters: "Go to my Instagram, DM me ‘SKILLS’ and I’ll send it over. Link’s in bio." At capture on 2026-10-19 05:09:20 UTC the metadata recorded 32,800 views, 2,235 likes, 122 comments, 473 reposts, and 2,708 saves. I read 10 of the 20 extracted frames (001, 002, 004, 006, 008, 010, 012, 014, 017, 020) and all 160 words of the transcript, 18 Whisper segments ending at 40.70 seconds. Frames 006, 008, 010, 014, 017, and 020 show one speaker at a desk in a black tee printed "AIS", with a Shure microphone, a gold YouTube play-button plaque, a white neon "AI" sign, and a book spine reading "AI NATIVE" and "NATE HERK"; in frame 020 he wears a black quarter-zip and points both index fingers up at the teal caption "SKILLS" in quotation marks. Frames 002 and 017 show a dark GitHub README table with columns "Skill", "What It Does", "Use When" under headings "Plan - Break it down", "Build - Write the code", "Verify - Prove it works", "Review - Quality gates before merge", and "Ship - Deploy with confidence". Frames 001, 004, 006, 010, and 012 are screen recordings of unlabeled coding-agent interfaces, detailed in the B-roll section below. Word-synced captions read "AN ENTIRE SOFTWARE" (001), "INTO 24" (002), "YOUR CODING" (004), "WORKFLOWS" (006), "HERE'S ONE example" (008), "CHANGING CODE" (010), "IT FIXES" (012), "AND THAT'S JUST" (014), and "DEVELOPMENT" (017). The transcript names "Addy Osmani's agent skills repo" at 4.22s, walks through a performance-optimization example from 15.06s to 26.74s, and ends with a DM call to action at 37.70s.

"Someone just turned" describes a repo created eight months before the post

GitHub's API records addyosmani/agent-skills as created on 2026-02-15 at 20:20:26 UTC, with its default branch most recently pushed on 2026-10-13 at 20:24:07 UTC, three days before this TikTok went up. The newest commit message on that push is "chore(release): bump plugin manifests to 0.6.13". The commit pagination header reports 702 commits. At my check on 2026-10-19 the repository showed 108,523 stars and 11,309 forks.

The video's opening line at 0.00s, "Someone just turned an entire software engineering playbook into 24 free AI skills", carries two framing choices worth separating. The first is "just". The repo is eight months old at post time and had already passed 27,000 stars when the author wrote about it on his own blog on May 3, 2026. The second is "someone". The transcript names him at 4.22s, "It's Addy Osmani's agent skills repo", and the spelling matches his GitHub handle. What the video never says is where he works. His GitHub bio reads "Member of Technical Staff, Anthropic", and his May 2026 post describes him as a Member of Technical Staff at Anthropic working on Claude Code after more than 14 years at Google. Several search snippets still call him a "Google Chrome engineer", and his GitHub profile's company field still reads "Google", so secondary coverage is stale on this point. The author's day job is relevant context for a repo that ships a .claude-plugin directory, and the video leaves it out.

The "24" matches the README's lifecycle tally, and the install pulls 25

The README I fetched on 2026-10-19 has a heading at line 224 reading "## All 25 Skills". The sentence under it says the pack includes "25 skills total" and breaks that down as "24 lifecycle skills plus the using-agent-skills meta-skill". The quick-start install at line 51 is npx skills add addyosmani/agent-skills with the trailing comment "install all 25 skills", and the GitHub contents API returns 25 directories under skills/. So the video's "24" is the README's own lifecycle count, and the number a user actually installs is 25.

The count has moved over the repo's life, which explains why older coverage disagrees. I pulled the earliest commit touching each skill directory. Eighteen directories, including performance-optimization, security-and-hardening, and test-driven-development, date to launch day, 2026-02-15. Seven arrived later: code-simplification on 2026-03-06, deprecation-and-migration on 2026-03-28, source-driven-development on 2026-04-08, doubt-driven-development on 2026-05-07, interview-me on 2026-05-10, observability-and-instrumentation on 2026-06-09, and constraint-driven-development on 2026-08-08. The 24-lifecycle figure has therefore been accurate since August 8, 2026, about ten weeks before the video. Earlier coverage reflects this: a DEV Community article title says "19 Production-Grade Skills", and the author's May 3 post says "twenty skills".

Frames 002 and 017 show the README's table with the headings "Plan - Break it down", "Build - Write the code", "Verify - Prove it works", "Review - Quality gates before merge", and "Ship - Deploy with confidence", which match README lines 243 through 277. Two headings sit above the crop in both frames: "Meta - Discover which skill applies" at line 228 and "Define - Clarify what to build" at line 234. Every skill name legible in frame 002, from interview-me down to performance-optimization, appears in the live table.

The performance example is a faithful paraphrase of the skill file, minus its fifth step

From 15.06s to 26.74s the video says: "Give your agent the performance optimization skill and a slow website. Instead of randomly changing code and hoping it just gets faster, it measures the site first, it finds the actual bottleneck, it fixes the specific problem, and then tests it again to prove that the change actually worked."

I fetched skills/performance-optimization/SKILL.md from the main branch on 2026-10-19. Its frontmatter name is performance-optimization and its description begins "Optimizes application performance across frontend, backend, queries, and databases. Use when performance requirements exist, when you suspect performance regressions, when Core Web Vitals or load times need improvement". The overview at line 10 opens "Measure before optimizing. Performance work without measurement is guessing" and later in the same paragraph says "Profile first, identify the actual bottleneck, fix it, measure again." The workflow block at lines 33 to 37 lists five steps, each label joined to its summary by an arrow: "1. MEASURE" and "Establish baseline with real data"; "2. IDENTIFY" and "Find the actual bottleneck (not assumed)"; "3. FIX" and "Address the specific bottleneck"; "4. VERIFY" and "Measure again; keep or revert"; "5. GUARD" and "Add monitoring or tests to prevent regression".

The video's four clauses map onto steps 1 through 4 in order, and "the actual bottleneck" is the file's own phrase. The video omits step 5, and it understates step 4. The file's "Step 4: Verify (Keep or Revert)" section says "A fix is a hypothesis until you re-measure", requires re-measuring "the way you measured the baseline: same command, same conditions, same fixed budget", and rules that "'Neutral' is a revert, not a keep." The frames never show this skill running; the "CHANGING CODE" caption at frame 010 sits over a generic React class component unrelated to measurement.

"Free" holds up: MIT license, an npx installer, and a format Anthropic calls an open standard

README line 410 reads "MIT - use these skills in your projects, teams, and tools." and the GitHub API reports the license as MIT. The README lists support for Claude Code, Cursor, Codex, Gemini CLI, Antigravity CLI, Windsurf, OpenCode, GitHub Copilot, Kiro IDE and CLI, and Command Code.

The "skills" in the video's title refer to a specific file format. Anthropic's engineering post "Equipping agents for the real world with Agent Skills", dated October 16, 2025, defines skills as "organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks", and says each skill is a directory containing a SKILL.md file that "must start with YAML frontmatter that contains some required metadata: name and description." The same post records that Anthropic published Agent Skills "as an open standard for cross-platform portability" on December 18, 2025. The live Claude Platform docs page I checked on 2026-10-19 lists the field rules and the Claude Code locations ~/.claude/skills/ for personal skills and .claude/skills/ for project skills. The format predates the repo by four months; the repo is one large pack written to it.

One line from those docs belongs next to the video's install pitch. The overview says "Use Skills only from trusted sources" and advises readers to "Treat like installing software", because a skill's instructions and scripts run with whatever access the agent has. Installing this pack drops 25 instruction files plus hooks and commands into your agent's path. "Free" and "safe to install blind" are different claims, and the video only makes the first.

The B-roll is unlabeled agent footage that never shows the repo in use

Frames 001, 004, 006, 010, and 012 are screen recordings from at least three visually distinct agent interfaces, none labeled on screen. Frame 001 is a light-themed IDE with a Swift project, CloudTicTacToe, and a change counter of "+389 -19". Frame 004 is a dark transcript where an agent reads KanbanDetail.tsx, edits a file whose line 26 reads **/.claude/settings.local.json, and adds # Playwright traces beneath it, before a todo list with three struck-through items and an unchecked "Build and test". Frame 006 is a browser-driving agent whose "Goal" block asks it to "Navigate to http://localhost:3000. Verify the dark mode aesthetic, the search bar, the album gallery, and the player bar", followed by "Typed 'M83' in Browser". Frame 012 is a light view headed "Fixing..." with seven unchecked webhook tasks, including "Implement fix with bounded exponential backoff" and "Create comprehensive tests", above tool calls "Task Reading HostViewController.swift", "Glob */webhook", "Grep webhool", and "Grep timeout|TIMEOUT".

The .claude/settings.local.json path in frame 004 is a Claude Code convention, but I cannot attribute any of these clips to a named product from the frames alone, and none of them shows the agent-skills repo installed or a skill being invoked. The footage illustrates "coding agent" generically while the audio describes one specific skill.

The DM mechanic, the creator, and one transcript artifact

The call to action is the only place the caption and the audio meet. The description reads "Go to my Instagram, DM me ‘SKILLS’ and I’ll send it over. Link’s in bio." and the speaker says at 37.70s "So just DM me this word on screen right here. If you're on TikTok, links in the bio." while frame 020 shows "SKILLS" in teal. The repo is public at github.com/addyosmani/agent-skills, so nothing in the video requires the DM to obtain the skills. I could not inspect what the Instagram reply contains, so what "it" refers to is unverified.

On the creator, his site's about page describes Nate Herk as based in Chicago, a former "Business Intelligence Analyst" at Goldman Sachs who co-founded TrueHorizon AI in January 2025, scaled it to "$100,000 per month in revenue", and exited in December 2025. It names his community "AI Automation Society" with "over 450,000 people", claims "more than 1,000,000 subscribers in under two years" on YouTube, and calls his book "Becoming AI Native" a "#1 Amazon bestseller". The "AIS" on his shirt is consistent with that community name, though the page never expands the acronym. The book spine in frame 006 matches the Amazon listing title "Becoming AI Native: Use AI to Get Real Results Without the Busywork".

Whisper rendered the segment at 33.88s as "So you're not just giving your AMR context". "AMR" is a transcription error of unclear origin, so I am not treating it as a quote. The hallucination check is otherwise clean: the last segment ends at 40.70 seconds against 40.93 seconds of audio.

Key Takeaways

  • Verified: The artifact facts are confirmed by ffprobe and the post metadata. 40.85 seconds, 1080x1920, HEVC Main at 60 fps, 3,705,862 bytes, posted 2026-10-16 15:04:48 UTC, 32,800 views and 2,235 likes at capture on 2026-10-19.
  • Verified: addyosmani/agent-skills exists under an MIT license, holds 25 skill directories, and showed 108,523 stars, 11,309 forks, and 702 commits when I checked on 2026-10-19.
  • Verified: The video's description of the performance skill matches the live SKILL.md, which prescribes MEASURE, IDENTIFY, FIX, VERIFY in that order and uses the phrase "the actual bottleneck". The video omits the file's fifth step, GUARD.
  • Correction: The video says "Someone just turned" the playbook into skills. GitHub records the repo created on 2026-02-15, eight months before the 2026-10-16 post, with a 0.6.13 manifest release on 2026-10-13.
  • Partial correction: The video says "24 free AI skills". The README says "25 skills total", described as "24 lifecycle skills plus the using-agent-skills meta-skill". The 24-lifecycle count has been accurate since 2026-08-08.
  • Clarification: The video says "an entire software engineering playbook". The README cites "Software Engineering at Google" and Google's engineering practices guide as sources; neither the README nor the author's blog post uses "playbook" for the repo.
  • Clarification: The video never states the author's employer. His GitHub bio reads "Member of Technical Staff, Anthropic"; search snippets calling him a "Google Chrome engineer" reflect his prior role.
  • Unverified: What the "SKILLS" Instagram DM sends. The repo is public without it.
  • Context: Anthropic's docs tell users to "Use Skills only from trusted sources" and to treat a skill like installed software. The SKILL.md format dates to Anthropic's October 16, 2025 post and was published as an open standard on December 18, 2025.

Resources

Published October 16, 2026. Writeup generated from a favorited TikTok.