The video calls Claudex "a new free skill" and gates it behind an Instagram DM; the project is a four-month-old MIT-licensed GitHub repo with 3,133 stars, and the video reverses which model builds and which one audits

The artifact is a 34.218 s MP4 (QuickTime mov,mp4,m4a,3gp,3g2,mj2 container, yt-dlp format id bytevc1_1080p_695018-1), 1080x1920 portrait, H.265/HEVC Main profile level 150 in yuv420p at a constant 60/1 fps across 2,052 frames with a 621,849 bps video bitrate, paired with an AAC HE-AACv2 stereo track at 44,100 Hz and 64,382 bps over 741 frames, for an overall 695,038 bps and a filesize of 2,972,855 bytes; ffprobe reports the video stream at 34.200 s and the audio stream at 34.218 s, and the separately extracted audio.mp3 at 34.299 s. The post timestamp is epoch 1791731009, which is 2026-10-11T15:03:29 UTC, from handle @nateherkai (uploader id 7621673628270494733, channel name nateherkai), with the audio credited as "original sound - nateherkai". The description is 73 characters: "Go to my Instagram, DM me 'CLAUDEX' and I'll send it over. Link's in bio." At capture time (metadata epoch field 1792387954, which is 2026-10-19T05:32:34 UTC, matching the metadata.json mtime of 2026-10-19 00:32:35 local) the counters read 24,800 views, 1,568 likes, 98 comments, 321 reposts, and 1,273 saves. I read all 17 extracted frames, frame_001 through frame_017, covering 0 s to 32 s at 2 s intervals, against a transcript of 13 segments and 144 words. Frame 001 (0 s) carries the captions "Free skill" in green and "THAT MAKES" in white over a split screen: the top half is a heavily blurred GitHub repository page whose README heading reads "Karpathy-Inspired Claude Code Guide" with the final word cut off, overlaid with a rounded app icon that is half orange burst and half OpenAI knot; the bottom half is the speaker in a dark "AIS" t-shirt at a Shure broadcast microphone, with a neon "AI" sign behind him, a YouTube play-button award plaque on a shelf, and a book spine reading "AI NATIVE" with the author line "NATE HERK". Frame 002 (2 s) shows a dark motion-graphic card headed PLAN.md with four monospace rows, scope inspect repository, plan lock requirements, build run proof checks, and review independent model, an orange burst mark top right and an OpenAI logo bottom right, under the caption "WORK ON THE". Frame 003 (4 s) is a full-bleed red card with the wordmark "CLAUDEX" above the split orange-and-black icon. Frame 004 (6 s), captioned "TO PLAN", shows a Claude chat response reading "Absolutely! Here's a detailed plan to build a fitness tracking app, covering the key phases, features, technology, and next steps." with an artifact titled "Fitness Tracking App Plan". Frame 005 (8 s) is a mockup with a black "TASK" pill above the split icon labelled "ROUTER", captioned "IT'S SPLIT". Frame 006 (10 s), captioned "FIRST", shows a card headed CLAUDE.md containing "1. Project Overview" and a green check reading "Clear overview / Easy to navigate, easy to maintain". Frame 007 (12 s) is the most informative screen in the video: a ChatGPT window showing a competitor advertising research table with rows for "Jimmy Joy", "Soylent", and "yfood", text about "Meta/Instagram placements" and "one Huel product", the figures "33% lower CPA" and "5M+ impressions and 50+ assets from 10 creators", the footer "ChatGPT can make mistakes. Check important info.", and a composer reading "Work with ChatGPT" with the model selector set to "GPT-6 Astra Medium", under the caption "THEN GPT-6 ASTRA". Frame 008 (14 s) is a talking-head shot captioned "ACTUALLY HOLDS UP". Frame 009 (16 s) is a motion graphic of three icons on a dashed loop. Frame 010 (18 s), captioned "AND CLAUDE AUDITS", shows a Claude Code style "Conversations" panel with a Bash entry reading IN npm run build / Test build and OUT Build succeeded with 26 warning cut off, then Create (tests/slack_webook.tsx) with diff lines 26 **/.claude/settings.local.json and 27 + # Playwright traces. Frame 011 (20 s), captioned "1 AI guide", is a mockup window with rows interface, workflow, proof, and ready. Frames 012 (22 s) and 013 (24 s) reuse the PLAN.md card under "2 FRONTIER" and "CHALLENGING / each other". Frame 014 (26 s) is the split icon under "RUNS AS A SKILL". Frame 015 (28 s) is the Codex wordmark under "OR CODEX". Frame 016 (30 s) is B-roll of the speaker typing at a laptop, captioned "WHENEVER / you build". Frame 017 (32 s) is a talking-head shot captioned "“CLAUDEX”". The spoken argument is that a free skill named Claudex splits a coding task across two frontier models so that one writes the plan and the other audits it, then the two swap roles for the build, and that the viewer should DM the word CLAUDEX on Instagram to receive it. No frame I read shows the skill installed, invoked, or running: there is no repository URL, no install command, no /claudex-loop or $claudex-loop invocation, no PLAN-REVIEW-LOG.md, and no terminal output from a cross-model review. The transcript's last segment ends at 34.06 s against an audio.mp3 duration of 34.299 s, so nothing is timestamped past the end of the audio and no Whisper hallucination is present; the transcriber did mangle two things inside the audio range, rendering "GPT-6 Astra" as "GBD6 Astra" in segments 1, 4 and 7, and rendering "grade its own homework" as "guide its own homework" at 19.70 s.

"There's a new free skill": the repository is four months old at post time and has been renamed twice

At 0.00 s the video opens with "There's a new free skill that makes Claude and GBD6 Astra work on the same task together." The transcription error is resolved by frame 007, where the ChatGPT model selector reads "GPT-6 Astra Medium", and by the burned-in caption at 12 s, "THEN GPT-6 ASTRA". The intended claim is Claude plus GPT-6 Astra.

The named project resolves to chaseai-yt/claudex-loop on GitHub. The GitHub API reports created_at of 2026-06-05T00:51:06Z, 3,133 stars, 340 forks, and a last push of 2026-10-14T17:20:17Z. The video posted on 2026-10-11, which puts the repository at four months and six days old on the day the word "new" was used about it. The repository's own README adds more history under "History and credits": "This repository was previously grill-me-codex and crucible; GitHub redirects still work. The original skills remain in [legacy/]." A project that has already been through two renames and shipped a legacy directory is not a fresh release.

The "free" half of the claim is accurate and then some. The README carries an MIT license badge and the credits line "[Claude Code Masterclass and Chase AI+] · [MIT license]". Anyone can clone it without a DM, without an Instagram follow, and without being on any list.

The singular also needs correcting. The README's table lists four skills: claudex-route for "A model recommendation or one scoped handoff", claudex-loop for "Requirements, plan review, implementation, and final inspection", codex-review, and codex-build. The README states plainly that "Claudex Route is independent of the Claudex Loop workflow".

The model pairing is configurable, and the repository's own text treats GPT-6 Astra as one option rather than the fixed partner

The video's framing from 0.00 s onward is that the skill pairs Claude with GPT-6 Astra specifically. The repository frames it as a pairing of two command-line tools, with the models left to the user's configuration. The README says: "Model choices remain configurable. Use Claude Fable 5.1 and GPT-6 Astra when selected and available on your accounts, or retain each CLI's configured model. The host UI selection does not automatically change the other CLI's configuration. Requested and observed model information is recorded separately, and there is no silent model/provider fallback."

Both model names are real. The Claude models overview lists Claude Fable 5.1 with the API id claude-fable-5-1, a 1M token context window, 128K max output, a June 2026 knowledge cutoff, and pricing of $10 per input MTok and $50 per output MTok. GPT-6 Astra was announced on 2026-09-03 with the API id gpt-6-astra, per OpenAI's own developer community announcement, which opens with "Anything you can do on a computer, Astra can do for you. Fast." and describes a phased rollout to Pro, Enterprise and Business Premium users in ChatGPT Work and Codex with API access live at announcement.

What the video omits is that the pairing is a parameter. The README's controls table lists reviewer_model, builder_model, and inspector_model, each defaulting to "each CLI's configuration" and described as "Explicit per-role model override". Its examples include claudex this feature, builder=codex, reviewer_model=gpt-6-astra and claudex this feature, builder=claude, reviewer_model=claude-fable-5-1. The claudex-route section names further candidates: "Luna may suit a focused fixture-generation task, Sol a bounded implementation, and Astra or Fable a difficult review." GPT-6 Sol and GPT-6 Luna were released on 2026-09-22 according to the GPT-6 Astra Wikipedia entry. A maintained downstream by ujconsulting lists gpt-6.1-sol/medium as the default model for reviewing roles, which is not Astra at all.

Who plans and who builds depends on which tool you start in, and the video reverses the Claude Code default

At 10.72 s the video says "First, Claude writes the plan, then GBD6 Astra audits it." At 17.20 s it says "Astra builds the project, and Claude audits the implementation." Both sentences are stated as fixed facts about the skill. The repository's README contradicts the second one for the most common entry point and qualifies the first.

The README's routing table reads, with columns for start point, requirements and plan, plan review, default build, and final inspection:

Claude Code: current Claude session, then Codex, then Claude, then a fresh Codex session.
Codex: current Codex session, then Claude, then Codex, then a fresh Claude session.

Two things follow. First, Claude only writes the plan when you start the loop in Claude Code. Start in Codex and the roles invert completely, with Codex planning and Claude reviewing. Second, and this is the direct correction, the default builder when you start in Claude Code is Claude, not the OpenAI side. The video's "Astra builds the project, and Claude audits the implementation" describes the Codex-host path, or a Claude Code run where the user has explicitly overridden the builder. The README makes the override explicit: "Choose either builder with builder=claude or builder=codex. The inspector follows the builder choice and always uses the other provider."

The underlying principle the video is reaching for does hold, and the repository states it in its own GitHub description: "Whoever built it never grades it." The README's build section says the final changes are inspected "with the other provider in a fresh session", and adds a condition the video never mentions: "If the coordinator takes over fixes, those new edits need another independent inspection."

"They keep going until the plan actually holds up" describes a bounded loop with an explicit round cap

At 13.50 s the video says "They keep going until the plan actually holds up", and at 22.16 s it says "You've basically got two frontier models constantly challenging each other before anything is finished." Both sentences imply the loop runs until convergence.

The repository caps it. The controls table sets rounds / MAX_ROUNDS to a default of 5, described as "Completed plan-review round cap", MAX_FIX_ROUNDS to 2 for "Build-fix attempt cap", and MAX_INSPECTION_ROUNDS to 2 for "Initial inspection plus one reinspection". Phase 3 says the reviewer "returns evidence-backed findings, and revisits revisions in the same session. Stop at the round budget or an explicit verdict: APPROVED, REVISE, or BLOCKED." A run can end at the budget with the plan unresolved, and the README treats that as a reportable outcome rather than a pass: "BLOCKED, execution failures and exhausted round budgets are surfaced rather than converted to approval."

The README is also more cautious than the video about what a passing review proves. Its section "What an approval means" states: "A clean structured result does not prove the model is right. The log preserves coverage, limitations and concrete evidence. Zero findings is valid; a large number of findings is not a quality score." It applies the same caution to its one published result: "The first reported end-to-end CRM planning run produced 55 findings over five rounds; it is an illustrative run, not a controlled benchmark of model pairings." The video quotes no figures, which avoids repeating an error, but it presents the adversarial loop as settled practice where the authors label their own evidence as illustrative.

"Drop it straight into Claude or Codex" skips two skill directories, two authenticated CLIs, and a Python runtime

At 26.34 s the video says "And because it runs as a skill, you can drop it straight into Claude or Codex and reuse the workflow whenever you build."

The format claim is correct. Claude Code's skills documentation defines a skill as a SKILL.md file with YAML frontmatter and markdown instructions, loaded from ~/.claude/skills/<skill-name>/SKILL.md for personal scope or .claude/skills/<skill-name>/SKILL.md for project scope, and states that Claude Code follows the open Agent Skills standard. OpenAI's Codex skills documentation confirms the same format, with personal skills in $HOME/.agents/skills and repository skills in $REPO_ROOT/.agents/skills, invoked by typing $ or running /skills. The agentskills.io overview describes the format as "a folder containing a SKILL.md file" and lists both Claude Code and ChatGPT with Codex among supporting clients.

"Drop it straight in" still understates the setup. The two tools read from different directories, so the README's install block copies the skills twice:

mkdir -p ~/.agents/skills ~/.claude/skills
cp -R skills/. ~/.agents/skills/
cp -R skills/. ~/.claude/skills/

The README's install section also names prerequisites the video does not: "Both CLIs must be installed and authenticated for the full cross-provider workflow. Python 3.10+ runs the shared adapter; no runtime pip packages or separate API keys are required. Check codex --version, codex login status, claude --version and claude auth status." It further warns that copying a single alias is not enough: "The compatibility commands share the runtime inside claudex-loop; copying an alias alone is insufficient." A viewer who installs only the Claude Code half and never authenticates the Codex CLI gets a skill that cannot perform the cross-model review the whole video is about. The video is correct that the result is reusable across projects, since both personal skill directories are machine-wide.

The deliverable is gated behind a DM, and nothing on screen is a recording of the skill running

At 31.06 s the video says "So just DM me this word on screen right here", and at 32.72 s "If you're on TikTok, link's in the bio." The word on screen at 32 s is "“CLAUDEX”". The description repeats the instruction: "Go to my Instagram, DM me 'CLAUDEX' and I'll send it over. Link's in bio." The funnel is the point of the video. The engagement numbers suggest it is working as intended: 1,273 saves and 321 reposts against 24,800 views at capture, with saves running higher than comments by a factor of roughly thirteen, which is the shape of a bookmark-and-come-back-later asset rather than a discussion.

Set against that, the gated item is MIT-licensed public code. The repository is indexable by search, carries 3,133 stars, and publishes its own one-line install path, /plugin marketplace add chaseai-yt/claudex-loop followed by /plugin install claudex-loop@claudex-loop. The DM adds a lead-capture step to an artifact that requires no permission to obtain. I cannot verify what the DM auto-responder actually sends, since that requires completing the funnel, so I make no claim about the destination beyond noting that the keyword matches a freely available project of the same name.

The on-screen material deserves the same scrutiny. The video implies a working system across its full 34 seconds and demonstrates it in none of them. The PLAN.md card at frames 002, 012 and 013, the CLAUDE.md card at frame 006, the TASK/ROUTER graphic at frame 005, the dashed-loop icon trio at frame 009, and the interface/workflow/proof/ready window at frame 011 are motion-graphic mockups with no file tree, no terminal, no cursor and no scroll state. The two real screen captures show other work. Frame 007's ChatGPT window is a nutrition advertising analysis naming Jimmy Joy, Soylent, yfood and Huel; its only contribution to the argument is the "GPT-6 Astra Medium" label in the composer. Frame 010's Claude Code panel shows npm run build returning "Build succeeded with 26 warning" and a file being created at tests/slack_webook.tsx, which is ordinary agent coding footage with no reviewer, no verdict and no second provider present. The opening B-roll in frame 001 is a GitHub page whose README heading reads "Karpathy-Inspired Claude Code Guide", matching the Karpathy-guidelines family of skill repositories rather than claudex-loop, whose README opens with a logo image and the line "### Skills for working across Claude Code and Codex."

Credit is the last gap. The repository names its authors: "Claudex Loop, cross-model review and packaging: [Chase AI]", with "Legacy interview skills: © [Matt Pocock]", and "Codex-as-builder pattern adapted from [Peter Steinberger]". The video names none of them and routes the audience to a personal Instagram DM instead of to the project. It does not claim authorship either, so this is an omission rather than a false statement.

Key Takeaways

Correction: At 0.00 s the video calls Claudex "a new free skill". GitHub reports chaseai-yt/claudex-loop created at 2026-06-05T00:51:06Z, four months before the 2026-10-11 post, with 3,133 stars and 340 forks, and its README says it "was previously grill-me-codex and crucible".
Correction: At 17.20 s the video says "Astra builds the project, and Claude audits the implementation." The README's routing table sets the default build to Claude when the loop starts in Claude Code, with Codex doing the final inspection. The OpenAI side builds by default only when you start in Codex, or when you pass builder=codex.
Partial correction: At 10.72 s the video says "First, Claude writes the plan." That holds only for a Claude Code host. Starting in Codex puts Codex on requirements and plan, with Claude reviewing.
Partial correction: At 13.50 s the video says the models "keep going until the plan actually holds up." The loop is capped: rounds defaults to 5, MAX_FIX_ROUNDS to 2, MAX_INSPECTION_ROUNDS to 2, and the README states that "exhausted round budgets are surfaced rather than converted to approval".
Clarification: The video presents Claude plus GPT-6 Astra as the fixed pairing. The README says "Model choices remain configurable", names Claude Fable 5.1 and GPT-6 Astra as a suggested pair "when selected and available on your accounts", and exposes reviewer_model, builder_model and inspector_model overrides. A maintained fork lists gpt-6.1-sol/medium as its default review model.
Clarification: The video says "a new free skill", singular. The repository ships four: claudex-route, claudex-loop, codex-review and codex-build, and states that Route "is independent of the Claudex Loop workflow".
Partial correction: At 26.34 s the video says you can "drop it straight into Claude or Codex". Both tools do read the same SKILL.md format, from ~/.claude/skills/ and $HOME/.agents/skills/ respectively, but the README requires both CLIs installed and authenticated plus Python 3.10+, and the install command copies the skills into both directories.
Verified: GPT-6 Astra is real. OpenAI announced it on 2026-09-03 with the API id gpt-6-astra, rolling out to Pro, Enterprise and Business Premium users in ChatGPT Work and Codex with API access live at announcement. Claude Fable 5.1 is real, with the API id claude-fable-5-1, a 1M token context window and a June 2026 knowledge cutoff.
Verified: The core design principle is stated by the project itself. Its GitHub description reads "Whoever built it never grades it", and the README specifies that the inspector "always uses the other provider".
Context: The transcript renders "GPT-6 Astra" as "GBD6 Astra" in three segments and "grade its own homework" as "guide its own homework" at 19.70 s. Frame 007's "GPT-6 Astra Medium" model selector and the caption "THEN GPT-6 ASTRA" at 12 s resolve the first; the repository's own "Whoever built it never grades it" resolves the second.
Context: No frame among the 17 I read shows Claudex installed or running. The PLAN.md, CLAUDE.md, TASK/ROUTER, loop-icon and interface/workflow/proof/ready screens are motion-graphic mockups. The two real captures show other work: a ChatGPT nutrition-advertising analysis naming Jimmy Joy, Soylent, yfood and Huel, and a Claude Code session running npm run build and creating tests/slack_webook.tsx. The opening B-roll shows a GitHub README headed "Karpathy-Inspired Claude Code Guide", which belongs to a different repository family.
Context: The repository credits Chase AI for Claudex Loop, Matt Pocock for the legacy interview skills, and Peter Steinberger for the Codex-as-builder pattern. The video names none of them and does not claim authorship.
Context: At capture (2026-10-19T05:32:34 UTC) the video had 24,800 views, 1,568 likes, 98 comments, 321 reposts and 1,273 saves. Saves outnumber comments by roughly thirteen to one, the pattern of a bookmarked lead magnet.
Unverified: What the Instagram DM auto-responder actually delivers in response to "CLAUDEX". Confirming it requires completing the funnel, so I make no claim about the destination.

Resources

chaseai-yt/claudex-loop, GitHub, created 2026-06-05, last push 2026-10-14 establishes the four-skill structure, the host routing table, the configurable model choices, the round caps, the install prerequisites, the MIT license, the rename history from grill-me-codex and crucible, and the credits to Chase AI, Matt Pocock and Peter Steinberger that I cite throughout.
GitHub REST API repository record for chaseai-yt/claudex-loop establishes the created_at of 2026-06-05T00:51:06Z, the 3,133 stars, the 340 forks, and the 2026-10-14T17:20:17Z last push that disprove "new".
ujconsulting/claudex-loop, GitHub establishes that a maintained downstream of the project defaults its review roles to gpt-6.1-sol/medium rather than GPT-6 Astra, and adds the read-only Codex wrapper and role gates.
Agent Skills, Claude Code documentation establishes the SKILL.md definition, the ~/.claude/skills/<skill-name>/SKILL.md and .claude/skills/<skill-name>/SKILL.md load paths, and the statement that Claude Code follows the open Agent Skills standard with Claude Code specific extensions that do not function in other tools.
Build skills, OpenAI Codex documentation establishes that Codex reads personal skills from $HOME/.agents/skills and repository skills from $REPO_ROOT/.agents/skills, and that skills are invoked with $ or /skills.
Agent Skills overview, agentskills.io establishes the open-standard folder format and the client list that includes both Claude Code and ChatGPT with Codex.
Models overview, Claude platform documentation establishes the current Claude lineup and the Claude Fable 5.1 facts I cite: API id claude-fable-5-1, 1M token context window, 128K max output, June 2026 knowledge cutoff, and $10 per input MTok with $50 per output MTok.
Introducing GPT-6-Astra, OpenAI Developer Community announcement, 2026-09-03 establishes the exact model name, the gpt-6-astra API id, the 2026-09-03 announcement date, and the phased rollout to ChatGPT Work and Codex with API access live.
GPT-6 Astra, Wikipedia establishes the 2026-09-03 limited preview and 2026-09-04 public release dates for Astra, and the 2026-09-22 release of GPT-6 Sol and GPT-6 Luna that the repository's routing section references.
Becoming AI Native, Amazon listing for Nate Herk establishes the title and author of the book visible on the shelf in frame 001, confirming the speaker's identity behind the @nateherkai handle.
swarmclawai/andrej-karpathy-skills, GitHub establishes that the "Karpathy-inspired coding-agent guidelines" README heading visible in the frame 001 B-roll belongs to a separate project family and not to claudex-loop.