<- all tokdocs
2026.10.23@networkchuck

NetworkChuck's caption credits the Hermes "Curator" for writing a skill on camera; the Hermes repo shows the line on screen is printed by `background_review`, and the Curator's first pass is deferred a full 168 hours

AI
Watch on TikTok

View on TikTok ->

A 77.438005 s vertical clip at 1080x1920 (aspect 0.5625, SDR), delivered as TikTok format bytevc1_1080p_758750-1 in a QuickTime/MOV container (mov,mp4,m4a,3gp,3g2,mj2): HEVC video, Main profile, level 150, yuv420p, 30/1 fps, 2,323 frames, 687,891 bps, alongside AAC audio in HE-AACv2 profile, 44,100 Hz stereo, 1,671 frames, 64,145 bps, for an overall 758,760 bps and a 7,344,610 byte file, confirmed by ffprobe rather than taken from metadata.json. Posted at epoch 1792792982, which is 2026-10-23 22:03:02 UTC, by uploader networkchuck (channel nickname "NetworkChuck", uploader_id 6557659014469566470), with audio credited as "original sound - NetworkChuck" and a 331-character description. At capture time, the metadata.json mtime of 2026-10-25 18:05:41 local, the counters read 7,660 views, 455 likes, 2 comments, 16 reposts and 147 saves. I read sixteen frames: 001, 005, 008, 010, 012, 016, 020, 023, 026, 029, 031, 033, 036 and 038 from the 2 s set, plus 063 and 071 from the 1 s set, against a transcript of 259 words in 28 segments. Frame 001 (0 s) is Chuck to camera in his studio with the burned-in caption "THE MOST POWERFUL THING". Frame 005 (8 s) cuts to a Windows desktop running the Hermes terminal UI: a header reading "Nous Research · Messenger of the Digital Gods", an ASCII caduceus, the footer "gpt-5.5 · Nous Research" over "/root", a right-hand panel showing "0 tools", the prompt "Who are you?>" answered "I'm Ron Weasley - grown-up, mildly battle-s[carred] ... systems. What do you want to work on today?", a following user line "I am NetworkChuck (you can call me Chuck) I'...", and a status bar reading "( '·_·' ) reflecting… | gpt 5.5 | voice off" above "Ctrl+C to interrupt…". Frames 008, 010 and 012 are talking head with the captions "AND WE'RE GONNA GIVE", "SOME SKILLS" and "WATCH IT'S GONNA HAPPEN". Frame 016 (30 s) splits the screen: a terminal tab titled "root@srv732018: " with three bare `root@srv732018:#prompts, a hand-drawn yellow cloud with two arrows labelled "Hermes", and two logo cards, the Nous Research anime avatar and the "Twingate" wordmark. Frame 020 (38 s) catches agent output reading "...d, and the intended path into Hogwart[s] ... anor via Twingate. Sensible. Much bet..." above a half-typed prompt "...r access to my s[tudio]". Frame 023 (44 s) shows four terminal tabs, "root@srv731968: ~", "root@srv731968: ~/.hermes/sk[ills]", "root@srv732018: ~" and "gpt-5.5 ·", with the agent listing "Plain terms: Hogwarts runs: • UniFi networking • MikroTik networking • Ceph for NAS/storage • Home Assistant for automation", then a user turn "you are on a hostinger vps in the cloud. You...", a collapsed "Tool calls (1) ● Memory("+memory "The assistant runs on a H…"")", and the replies "Got it, Chuck - saved in the Pensieve." and "Plain terms: I'm running on a Hostinger VPS". Frame 026 (50 s) is the Twingate verification block: "Twingate status: online", "Twingate interface: sdwan0", "Hogwarts route present: 10.77.14.0/24 dev sdwan0", "Non-routable Twingate range present: 100.96.0.0/12 dev sdwan0", "I also did one harmless ICMP probe to 10.77.14.1. It didn't [respond] ... That host may simply not exist or may blo[ck ping]", a security note that "since the service key was pasted into cha[t]", and the offer "Want me to save this Twingate headless setup process as a re[usable spellbook]". Frames 029, 031, 033 and 1 s frame 063 are the payoff: the typed instruction "Search the network, see what you can find." followed by a yellow-circled line reading💾 Self-improvement review: skill "twingate-client-operations" created, with the status bar now "( °□ °) reasoning… | gpt 5.5 | voice off". Frame 036 (70 s) is a blurred code montage including "from hermes_tools import terminal import...", "set -euo pipefail printf 'Twingate commands:", "set -euo pipefail twingate resources || true", three Code("import socket...andCode("import urllib.request, ssl, re...` fragments, and a counter reading "ng 13/54…". Frame 038 (74 s) is a near-black outro card. The spoken argument is simple: Chuck gives his cloud-hosted Hermes agent, personified as Ron Weasley, a Twingate service key, the agent builds a headless Twingate client and reaches his home lab, and then without being asked it writes itself a reusable skill, which Chuck presents as the single most important thing Hermes does. On the transcript-end check, the last substantive segment ends at 74.72 s against an audio stream of 77.479184 s, so nothing of substance is hallucinated past the audio; the final six segments are bare "." punctuation spanning 77.24 s to 77.48 s, the last of which overshoots the audio end by 0.001 s, a rounding artifact carrying no quotable content.

The caption credits the Curator; the repo says the line on screen is printed by the background review

The video's own description states: "That's exactly what the Curator feature in Hermes does. It helps you refine and improve your skills, removing the ones that don't work and creating new ones automatically." Half of that is right and half is wrong, and the frames settle it.

The string circled in yellow in frames 031 and 033 is 💾 Self-improvement review: skill "twingate-client-operations" created. That string is a template in the Hermes repository. In locales/en.yaml the block is introduced by the comment "Self-improvement review summaries (agent/background_review.py)." and contains summary_cli: " 💾 Self-improvement review: {summary}" and summary_callback: "💾 Self-improvement review: {summary}", with the sub-template skill_result: "Skill '{name}' {verb}{path}" and skill_verb: created: "created". The emitting module named in the repo's own comment is agent/background_review.py, not the curator.

The Curator is a different subsystem. website/docs/user-guide/features/curator.md opens: "The curator is a background maintenance pass for agent-created skills. It tracks how often each skill is viewed, used, and patched, moves long-unused skills through active → stale → archived states, and periodically spawns a short auxiliary-model review that proposes consolidations or patches drift." The next paragraph draws the line explicitly: "It exists so that skills created via the self-improvement loop don't pile up forever. Every time the agent solves a novel problem and saves a skill, that skill lands in ~/.hermes/skills/."

So the Curator cleans up skills the self-improvement loop has already created. The description's phrase "removing the ones that don't work" describes the Curator accurately. The phrase "creating new ones automatically" describes the background review, which is what the video actually filmed.

The Curator's shipped defaults rule it out as the cause of what happens on camera

Even without the locale file, the timing in the video excludes the Curator. The same doc gives the default configuration block verbatim:

curator:
  enabled: true
  interval_hours: 168          # 7 days
  min_idle_hours: 2
  stale_after_days: 14
  archive_after_days: 30
  consolidate: false
  prune_builtins: false

Three of those values matter here. interval_hours: 168 is a seven-day cadence. min_idle_hours: 2 means the Curator only fires after the profile has been idle for two hours, and the doc adds that Hermes will "skip curator while a turn in that profile is running". In frames 026 through 033 a turn is visibly running: the status bar reads "reasoning…" and the footer offers "Ctrl+C to interrupt…". And the doc states that on a fresh install "the curator does not run immediately. The first observation seeds last_run_at to 'now' and defers the first real pass by one full interval_hours."

The third value is consolidate: false. The doc is unambiguous: "The opinionated LLM consolidation pass (umbrella-building, merging overlapping skills) is off by default because it costs aux-model tokens on every run and makes broad structural changes to your library." Consolidation is the only Curator pass that can call skill_manage at all, and it ships disabled.

The background review ships the other way round. cli-config.yaml.example documents it as the "Post-turn memory/skill self-improvement review fork. Runs after a turn when the nudge intervals fire; writes skills/memories in a daemon thread", with enabled: true and the comment "false = skip automatic forks (/refine still works)". A fork that runs after a turn, in a daemon thread, enabled by default, is exactly what produces a skill-creation line seconds after Chuck types "Search the network, see what you can find."

"It makes its own skills" is the correct claim, and it is Hermes' documented headline behaviour

At 71.08 s the video says "That's the power of Hermes. It makes its own skills." That holds up. The repository description is "The agent that grows with you", and the project homepage lists under Persistent Memory: "It learns your projects, auto-generates skills, and never forgets how it solved a problem."

website/docs/user-guide/features/skills.md sets out the mechanism. Skills live in ~/.hermes/skills/, "the primary directory and source of truth", which matches the second terminal tab visible in frame 023, titled "root@srv731968: ~/.hermes/sk[ills]". The agent writes them through a meta-tool called skill_manage with actions create, patch, delete, write_file and remove_file. The documented triggers are that the agent "worked out a multi-step workflow worth repeating", "hit errors or dead ends and found the working path", or that "the user corrected its approach". Chuck's Twingate setup is the first of those: a multi-step headless install worth repeating.

The doc also confirms the two writers are distinct and both real: a skill change can arrive "in a foreground turn, by the background review, or by the curator's consolidation pass". The write-approval gate, skills.write_approval, defaults to false, which is why nothing in the video pauses to ask Chuck's permission before the skill lands.

One detail the frames support independently: the name on screen is the hyphenated identifier twingate-client-operations, not the spaced phrase the burned-in caption renders as "TWINGATE CLIENT OPERATIONS CREATED". Frame 029 shows the quoted form "twingate-client-operations" created at native zoom.

Twingate sells itself as a VPN replacement, not a VPN

At 28.16 s the video says "We're gonna use twin game. It's a VPN. It's a way to remote into my stuff." Two things are wrong in that sentence, one of them Whisper's fault and one Chuck's compression.

The product name is Twingate, spelled as one word, confirmed by the logo card in frame 016 and by the agent's own output in frames 020, 026 and 033. Whisper rendered it "twin game" and "twin gate" throughout, which is the exact failure mode where a product name decays into plausible English.

On the category, twingate.com leads with "Security, Performance, Simplicity. Pick Three." and describes the product as "Identity-based access for users, services, and AI agents that deploys in minutes, scales to every resource, and finally lets you retire your VPN." Twingate positions itself as Zero Trust Network Access and markets against VPNs rather than as one. Calling it "a VPN" in a 77-second clip is defensible shorthand for the outcome, and it misstates what the vendor says it sells. The agent output in frame 026 is more precise than the voiceover: it reports a dedicated interface sdwan0, a specific route 10.77.14.0/24, and the Twingate-internal non-routable range 100.96.0.0/12, which is resource-scoped routing rather than a blanket tunnel.

"I didn't tell him to do this" is true of the instruction and not of the offer

At 60.8 s Chuck says "I didn't tell him to do this." Narrowly that is correct: the prompt visible in frames 031 and 063 is "Search the network, see what you can find.", which says nothing about skills.

The frames add a qualifier the voiceover skips. One turn earlier, in frames 026 and 031, the agent's own message ends with "Want me to save this Twingate headless setup process as a re[usable spellbook]". The agent had already proposed writing the skill and was waiting on an answer. Chuck typed an unrelated instruction instead, and the background review fork wrote the skill anyway on the next turn boundary. That sequence is more interesting than the one the voiceover describes, because it shows the fork acting independently of a pending foreground offer rather than inventing the idea from nothing. The on-screen record and the claim do not contradict each other, but the clip omits the half-second of context that explains where the idea came from.

The clip never shows the skill, and the frames correct three names the transcript mangles

Across all sixteen frames I read, the video never opens twingate-client-operations. There is no cat of its SKILL.md, no skills_list output, no file tree under ~/.hermes/skills/, and no second run that loads the skill to prove it works. The ~/.hermes/sk[ills] tab is visible in frame 023 but is never brought to the foreground. The evidence shown is one log line asserting creation. Whether the skill is correct, complete, or reusable is not demonstrated in this clip. The word "Curator" appears in none of the frames I read, which is notable given that the description names it as the feature on display.

Three names the transcript gets wrong, corrected from the frames and from primary sources. "twin game" and "twin gate" are Twingate. "He's in the hosting your cloud" at 23.6 s is Hostinger: frame 023 shows Chuck's typed line "you are on a hostinger vps in the cloud" and the agent's reply "Plain terms: I'm running on a Hostinger VPS", and Hostinger sells VPS hosting under that spelling. "our it agent" at 11.32 s is "our IT agent".

On the model, the status bar in frames 005, 023, 031, 033 and 036 reads "gpt 5.5" and the TUI footer in frame 005 reads "gpt-5.5 · Nous Research". That is OpenAI's GPT-5.5, which TechCrunch reported on 23 April 2026: "OpenAI on Thursday released GPT-5.5, its newest AI model, which the company calls its 'smartest and most intuitive to use model' yet." Hermes is model-agnostic; its homepage advertises access to "200+ cutting-edge models" via Nous Portal and the repo README offers "Nous Portal, OpenRouter, OpenAI, your own endpoint, and many others". Nothing in this demo depends on Nous Research's own Hermes model family, despite the shared name. The header string itself is real repo text: ui-tui/src/i18n/en/chatBits.ts defines tagFull: 'Nous Research · Messenger of the Digital Gods', which is verbatim what frame 005 shows.

Key Takeaways

  • Correction: The video description says "That's exactly what the Curator feature in Hermes does ... creating new ones automatically." The skill-creation line shown on screen is templated in locales/en.yaml under a comment naming agent/background_review.py, and curator.md states the Curator exists to maintain "skills created via the self-improvement loop", not to create them.
  • Correction: The Curator's shipped defaults make it impossible as the cause of the on-camera event: interval_hours: 168, min_idle_hours: 2, consolidate: false, and a documented rule that the first pass on a new install is deferred by one full interval. The background review ships enabled: true and runs after a turn in a daemon thread.
  • Partial correction: At 28.16 s the video calls Twingate "a VPN". Twingate's own site describes identity-based Zero Trust access that "finally lets you retire your VPN". The agent's output in frame 026 shows resource-scoped routing on sdwan0, not a blanket tunnel.
  • Verified: "It makes its own skills" at 71.08 s matches the documented behaviour. skills.md describes agent-written skills in ~/.hermes/skills/ via the skill_manage tool, with skills.write_approval defaulting to false, which is why no approval prompt appears.
  • Clarification: "I didn't tell him to do this" at 60.8 s is true of the typed instruction. Frames 026 and 031 show the agent had already asked "Want me to save this Twingate headless setup process as a re[usable spellbook]" one turn earlier and never got an answer.
  • Clarification: The on-screen skill identifier is twingate-client-operations (frame 029). The burned-in caption renders it as the spaced phrase "TWINGATE CLIENT OPERATIONS CREATED".
  • Unverified: Nothing in the sixteen frames I read opens the created skill. No SKILL.md contents, no skills_list, no second run exercising it. Creation is asserted by one log line and not demonstrated.
  • Context: Whisper mangled three names. "twin game"/"twin gate" is Twingate, "hosting your cloud" is Hostinger, and "our it agent" is "our IT agent". All three were resolved against the frames and the vendors' own sites.
  • Context: Hermes Agent is MIT-licensed and the repository was created on 2025-07-22. At the time I queried the GitHub API on 2026-10-26 it showed 257,220 stars and 56,346 forks, and the latest release was v0.22.5, published 2026-10-23T05:28:45Z, roughly seventeen hours before this clip went up.
  • Context: The demo runs on OpenAI's GPT-5.5, released 23 April 2026, through a Nous Research agent. The Hermes agent and the Hermes model family share a name and are not the same thing, and this clip uses none of the latter.

Resources

Published October 23, 2026. Writeup generated from a favorited TikTok.