← All sessionsHomeSearch
AI Sprints (Live Weekend Programs)·The Hermes Agent Sprint·2:28:07

AI Sprint: The Hermes Agent — Day 1 (Always-On Agents, the Employee-on-Day-One System, and the Self-Improving Ghostwriter)

Dileep (KVSS Dileep) Head of Generative AI Education at Outskill; teaches Hermes as a case study in designing self-correcting agent systems — deliberately unstructured ('I'll be debugging on the fly... I don't want to show you a trailer of MasterChef; I'll cook the food in front of you') · Kartik Sprint host — logistics, certificates form, recordings timing, polls
▶local recording  ·  LMS lesson

Session map

THE WHYTHE SYSTEMTHE CRAFTThe dynamo and the computerwhy bolted-on AI produces bills, not pr…Hermes ≡ OpenClaw ≡ (mostly)…the always-on agent categoryEmployee on day onesoul.md, user.md, memory.md — onboardin…The self-improving ghostwriterfeedback → named lesson → patched skillRunning the agent through an…Codex as installer, debugger, and escap…
The whyThe systemThe craft
click a node — its card pops up (drag it anywhere, × to close)
Concept

The map reads left to right — the why flow into the system, then into the craft. Click any node to open that idea here; every timestamp jumps into the recording.

The short version

  1. The framing lesson is a 1990 economics paper: 'The Dynamo and the Computer.' Factories that swapped steam engines for electric ones saw NO productivity gain — only electricity bills; the gains came when they redesigned the factory around decentralized electric motors. 'That's exactly what's happening today: people buy ChatGPT and Claude subscriptions, and what goes through the roof are AI bills.' Reimagine the process around AI; don't bolt AI onto the old process.
  2. Hermes defined in one line: a 24/7 always-on agent, 'very, very similar to OpenClaw' — and immediately demoted from subject to specimen: 'I am not teaching Hermes. Hermes is an excuse. The main thing I'm going to teach you is a system that can work for you' — because Hermes, Claude Code, and Codex 'have become more or less the same' and are used interchangeably (supervised work → Claude Code/Codex; always-on autonomy → Hermes; md files carry the handoff).
  3. The mental model that organizes setup is EMPLOYEE ON DAY ONE: AI's default is Dory-from-Finding-Nemo memory, so you write the onboarding docs — soul.md (who the employee is: his spec demands no sugarcoating, no sycophancy, pushback, analogies, no emoji, no em-dashes, a Confucius quote after every task), user.md (who the boss is and how they like reports), memory.md (project facts, tools, known bugs), then skills. Setup is an INTERVIEW: 'interview me so I can give you the insights.'
  4. The build is a self-improving LinkedIn ghostwriter: voice DNA extracted from his real posts → a skill → then the upgrade that changes everything: 'whenever you take user feedback, tell me what you learned, and with every lesson, patch the skill.' The loop catches him a fabricated example on the first draft ('I invented it... that is worse than having no example at all'), learns to demand real anecdotes and real research, learns the 3,000-character cap — 'the lessons-learned section starts empty and grows with every correction.'
  5. Power = access: 'It's exactly like asking, can my gardener cut the grass? The gardener can cut the grass if you give him the tool.' Scrape Creators, Firecrawl, parallel search, X search, session search — capability questions ('can Hermes do X?') are dismissed as tool-access questions, with $3-5 experiment budgets keeping the blast radius small.
  6. The meta-craft: he runs Hermes THROUGH Codex — installs it by pasting a GitHub URL, wires Ollama by asking, debugs Hermes' image-API failure from Codex, and when the bug resists, switches surfaces without drama and finishes the explainer-image skill in Codex — the interchangeability doctrine demonstrated by its own failure handling.

The concepts

01

The dynamo and the computer: why bolted-on AI produces bills, not productivity

Factories replaced their central steam engine with an electric one and got nothing. 'Their productivity did not improve... there were absolutely no accrued benefits.' Sound familiar?

The 1990 paper (surfaced to him via a Zara Zhang tweet he re-reads 'from time to time') describes electricity's adoption lag: swapping the central steam engine for a central electric engine changed nothing, because the factory was still shaped around a central engine. Productivity arrived only when factories were REDESIGNED — decentralized, one electric motor per mill. The AI translation is exact: 'companies get ChatGPT and Claude subscriptions for employees, then wait for productivity to go through the roof — but what goes through the roof are AI bills.' You have to 'reimagine AI as the center and build around it,' knowing the industry hasn't fully found the answer yet — 'the mechanism remains the same.'

The proof case is Outskill's own Viber YouTube channel: the limiting belief was that long videos need a physical person; the redesign made the SYSTEM central (one metric — average view duration; AI clones for presence; AI-taught research and taste), persisted through months of 'it's not Vaibhav, it's AI' comments, and inflected when Seedance v4 (in HeyGen) delivered a step-change in clone quality. 'Our reimagination was removing the limiting belief.'

Worked example · from the session

The steam-to-electric swap mapped one-to-one onto seat licenses: same workflow + new engine = same output, higher bills.

Why it matters

It's the sprint's why-bother: every md file and feedback loop that follows is 'redesigning the factory' instead of swapping the engine.

People get this wrong

Buying frontier-model subscriptions for staff is AI adoption.

That's engine-swapping. Adoption is process redesign: new workflows, new metrics, new divisions of labor with agents at the center — and it takes iterations, as electricity did.

You cannot replace steam with electricity and expect productivity. You have to take electricity as the center and rebuild the factory around it.
People are getting ChatGPT and Claude subscriptions... and what are going through the roof are AI bills.
For your projects

Your own KB pipeline is the redesigned-factory version of note-taking — worth using as the concrete example when you tell this story.

This idea elsewherecomplementsthe Marketing OS is a factory-redesign exhibit: tools rebuilt around one command center
Go deeper

In one line: Dynamo parallel = general-purpose technologies pay off only after process redesign around their new shape (decentralized, always-on, feedback-driven), not after in-place substitution; AI adoption that preserves the old workflow produces cost without productivity.

Redesign includes metrics: the YouTube case chose ONE number (average view duration) and subordinated everything to it ()

Persistence through the ugly middle is part of the redesign — the clone-hate comments era preceded the quality inflection ()

Model progress is a step function you position for: Seedance v4 changed the equation overnight for a system already built around clones ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What's the 'central steam engine' in a typical company's AI rollout?

The unchanged workflow — same meetings, same documents, same approvals — with AI licenses bolted on where a human used to type. The redesign question is which processes get rebuilt around an always-on agent instead.

02

Hermes ≡ OpenClaw ≡ (mostly) Claude Code/Codex: the always-on agent category

'A 24/7 always-on agent... and I will tell you something nobody will tell you: you can work interchangeably with any of this.'

The category: Hermes (Nous Research, MIT, open source) and OpenClaw are agents that run continuously — on your own machine (kept on) or a remote server (Hostinger sells 1-click installs for both) — reachable by text from anywhere, including Telegram. Hermes' differentiating claim is 'self-improving: the more you use it, the better it gets' — which Dileep immediately grounds: self-improvement is a SYSTEM you set up (feedback → lessons → patched skills), and 'Codex and Claude Code are not self-improving until you set up some systems to do that.' The working division of labor: supervised, interruptible work → Claude Code or Codex (also his token-overflow relief valve for each other); always-on autonomy and session-memory → Hermes; md files make the handoff portable in every direction.

Practical adulthood is sprinkled through: security = WHERE you install it ('if you're unsure, don't; if it's corporate, your IT team must know; proprietary data → local models'); cost = the harness's tokens ('installing via Codex, the price is Codex tokens'); model choice = economics ('why Opus 4.6 when 4.8 exists? Old models get cheaper. I'm okay with this'); and 'for the same plan, Codex gives you much more tokens than Hermes.'

Worked example · from the session

The install itself: paste the GitHub URL into Codex, 'install Hermes if it is not already installed' — 10-15 minutes, no tutorial followed; the manual path exists for those unwilling to spend tokens.

Why it matters

It's the calm, vendor-neutral map of the agent landscape the course's other sessions gesture at — harnesses converge; systems differentiate.

People get this wrong

Hermes is a different kind of intelligence than Claude Code or Codex.

Same models, different harness posture: always-on + session memory + self-improvement HOOKS. The intelligence differences you feel are configuration and access.

I am not teaching Hermes. Hermes is an excuse for me. The main thing I'm going to teach you is a system that can work for you.
For your projects

Your scheduled tasks are the always-on tier of this project; this concept names why they carry the full harness (queue, rules, method) in their prompts.

This idea elsewherecomplementsthe md-folder portability doctrine, taught from the agent sidebuilds onthe Catalyst harness theory this sprint operationalizes
Go deeper

In one line: Always-on agent = continuously running, remotely reachable agent process (Hermes/OpenClaw; local or 1-click VPS) distinguished from supervised CLI agents (Claude Code/Codex) by availability and session memory, not by capability; all interchangeable via md-file harnesses, with 'self-improvement' being configuration, not magic.

'The LLMs don't know what they are' doctrine from the Catalyst track reappears as practice: same principles run on any of the four harnesses ()

Folder-scoped like the others: cd into the project, type hermes — 'exactly like how you would say claude code' ()

Available tools out of the box: browser, computer use, skill packs (creative/data-science/DevOps/MCP/media) — permissioned by you ()

Hostinger's 1-click tier is the sandbox recommendation for the nervous: 'a safer sandbox to play with' than your own laptop ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

When does always-on beat supervised for a task?

When the task outlives your attention — scheduled work, long loops, inbound triggers (Telegram/WhatsApp). If you'd watch it run anyway, a supervised CLI agent is cheaper and easier to steer.

03

Employee on day one: soul.md, user.md, memory.md — onboarding docs for an agent

'Reframe the question. Not: how do I use Hermes effectively. But: how do I train and set up my new employee for success?' Then write the onboarding docs.

The memory problem first, in analogy triplicate: default AI is Dory from Finding Nemo / Memento / Ghajini — 'you spend three hours teaching Claude to write a good proposal, close the laptop, and tomorrow the same mistakes repeat.' The fix is a file architecture the agent reloads every session: SOUL.MD — who the employee is and how they behave (his dictated spec: 'no-nonsense... tells me things as it is without sugarcoating and absolutely no sycophancy; concise, direct; push back if I'm wrong; analogies wherever possible; think out loud; no apologies, no filler, no emoji, no em-dashes' — plus the one quirk: a task-relevant Confucius quote after every completion). USER.MD — who the boss is: 'if you don't know your boss, what they like, what formatting they want, your boss will always be pissed at you' — his covers role, daily work, directness preference, scannable output (bullets/tables/checklists). MEMORY.MD — project-level durable facts: tools, conventions, known bugs and workarounds. SKILLS are the fourth layer — 'if you want multiple employees, set them up as skills; don't multiply user/memory files.'

The setup method is the reusable part: 'INTERVIEW ME so I can give you the insights, and you write the file' — the agent asks, you dictate (Wispr Flow), it drafts, you correct ('you have asked the wrong questions' → 'you're right, my mistake — soul.md is about ME, not you'). Files stay compact: 'you cannot write essays over here.'

Worked example · from the session

Confucius fires immediately and aptly for the rest of the session — 'a man who has committed a mistake and does not correct it is committing another mistake' arriving precisely on the feedback-loop lesson.

Why it matters

It's the concrete answer to agent amnesia — the same architecture s20 discovered as agents.md+goal.md, here split by ROLE (agent/boss/project) instead of by function.

People get this wrong

Memory features in chat products solve this.

Product memory is opaque and non-portable; the md architecture is inspectable, editable, versioned, and moves between harnesses — the employee file, not the vendor's impression of you.

An employee on day one: the onboarding paperwork soul.md values · boundaries user.md who it works for memory.md what it learns Hermes the AI employee on Telegram Your day briefings · drafts while you sleep Same architecture as CLAUDE.md/memory.md and the OpenClaw documents — behavior and state, split on purpose
soul.md, user.md, memory.md: onboarding paperwork for a day-one employee
Reframe it: how do I train and set up my employee for success? What is the system? Implement that system. That's all it is.
If you don't know your boss, what they like, what formatting they like, your boss will always be pissed at you. Therefore, tell your employee what you like.
For your projects

His user.md preferences are nearly YOURS verbatim (direct, no fluff, pushback, scannable output) — reading this session's spec is like reading your own preference block.

This idea elsewherecomplementsagents.md+goal.md and soul/user/memory are the same doctrine, sliced differently
Go deeper

In one line: Agent onboarding = role-separated persistent context: soul.md (agent persona + behavioral contract), user.md (boss profile + reporting preferences), memory.md (project facts, tools, known bugs), skills (per-job capabilities); authored by interview, kept compact, reloaded every session.

The persona spec doubles as an anti-sycophancy contract — pushback and blind-spot surfacing are WRITTEN DUTIES, not hopes ()

Verification habit: ask the agent whether the file exists and what it says — his user.md turned out to live in the memory system's profile layer, found by asking ()

Global vs project scoping is explicit: persona and boss are global; memory.md belongs to the project ()

Cross-agent: the same interview produced files usable from Codex or Hermes interchangeably — the files ARE the employee ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What belongs in memory.md that must never be in soul.md?

Perishable project truth — tool configs, conventions, known bugs and their workarounds. Soul.md is durable persona; mixing them makes both stale.

04

The self-improving ghostwriter: feedback → named lesson → patched skill

'Whenever you take user feedback, tell me what you learned — and with every lesson, keep improving the skill.' One instruction turns a writer into a learner.

The build sequence: extract VOICE DNA from his real LinkedIn posts (attach the file, 'extract my voice DNA', convert the XML spec to markdown) → install as a LinkedIn-assistant skill → add the LOOP INSTRUCTION above. Then feedback does the teaching, on camera: draft one is 'stale' and its example FABRICATED — the agent names its own lessons ('Lesson 1: I invented the meeting-notes scenario to fill the concrete-example slot. That is worse than having no example at all. Lesson 2: ask for real examples. Lesson 3: no research was done'), patches the skill, and asks HIM for a real anecdote (the enterprise client who never told Claude how he thinks). Draft two lands with real research (the MIT 95%-of-AI-projects-fail stat) and his real story. Subsequent corrections accrete the same way: the 3,000-character LinkedIn cap becomes a permanent lesson; 'the lessons-learned section starts empty and grows with every correction — future posts benefit from every past one.'

The governance details that make it durable: an over-engineering release valve ('anytime the skill feels too constrained, say remove these points'); grounding tools wired in (X search, web extract, parallel search, and SESSION SEARCH over past conversations — the agent mines its own history for your preferences); source-weaving rules negotiated by interview (drop a URL → it asks intent, attribution, and your take — 'merge it into my own thought instead of plagiarizing'); and the honest definition when a viewer challenges 'this just looks like refinement': 'self-improving comes when it understands feedback and makes corrections — an employee gets the realization when the boss gives feedback.' Boris' line is invoked as the frame: 'I don't write prompts. I do loops.'

Worked example · from the session

The second skill built the same way: Jack Butcher's Visualize Value principles scraped into an explainer-image playbook (his md file turned out to be a WRITING system, so the design article was used instead), outskill.com palette via Firecrawl's cached scrape, images through OpenRouter — one working illustration by session end despite a screenshot-pixel API error.

Why it matters

This is the sprint's deliverable: not a LinkedIn tool but the loop template — any skill + the feedback-to-lesson instruction = an asset that compounds.

People get this wrong

Self-improvement means the agent gets better on its own.

It means feedback is CONSERVED: every correction becomes a named, persisted rule. No boss feedback, no improvement — on Hermes or anywhere else.

I fabricated an example... I invented it to fill the concrete example slot. That is worse than having no example at all.
The lessons-learned section starts empty and grows with every correction. Future posts benefit from every past correction.
Self-improving comes when it understands feedback and makes corrections... the employee gets the realization when the boss gives the feedback.
For your projects

This is the strongest single argument in the corpus for your skill-update habit — the 'stated lesson' rule is directly adoptable into how your skills get amended.

This idea elsewherebuilds onthe Catalyst loop anatomy, embodied in a shipping skillcomplementsthe code-side loop; this is the content-side twin
Go deeper

In one line: Self-improving skill = capability spec (voice DNA / design principles) + explicit loop contract (on feedback: state the lesson, patch the skill, rewrite) + grounding tools (web/X/session search) + release valve against over-constraint; improvement is boss-fed, named, and persisted — never assumed.

Hallucination handled as a curriculum moment: the fabricated example becomes the skill's first permanent rule ()

Session search is the underrated tool: the agent reads its own past sessions to learn how you write and decide ()

Platform constraints are lessons too — the 3,000-char cap patched in after one violation ()

The loop ran even where he didn't write the prompt: the final image's generation prompt was composed entirely by the system — 'that is the loop' ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why must the agent STATE the lesson before patching the skill?

The stated lesson is the audit trail — you verify the generalization is right before it becomes permanent; silent patching accumulates wrong rules you can't see.

05

Running the agent through an agent: Codex as installer, debugger, and escape hatch

'When I face an error in Hermes, I don't debug in Hermes. I bring it into Codex — my Codex can drop to the Hermes CLI.'

The operating pattern: Codex (or Claude Code) is the MANAGEMENT layer for every other tool. Install Hermes? Paste the GitHub URL: 'install Hermes if it's not already installed.' Connect local models? 'I want to connect Hermes to Ollama — how?' then 'I am non-technical, just set this up on my behalf' (Codex checks the 16GB RAM, declines the 32B model, pulls a small coder, edits Hermes' config). Understand the file architecture? Ask Codex to run a parallel web search on soul.md vs user.md best practices, then use the answer to configure Hermes. Debug? The image-generation failure (a full-page screenshot exceeding pixel limits) is diagnosed FROM Codex; when the fix resists stage time, he switches surfaces entirely — 'some problem with Hermes; I can debug it but it'll take time' — and finishes the skill in Codex without losing the thread. 'What I'm doing is using Codex as a harness for Hermes.'

The supporting discipline: watch the run and steer early ('wherever it deviates, I stop it — if I don't, it keeps diverting'); give tools access or accept impotence ('you want power, you need to give access... the gardener can cut the grass if you give him the tool'); cap the blast radius ('I only put $3 to $5 into my tools — maximum I lose is $3'); and answer capability questions yourself ('if your question starts with does Hermes— , ask Perplexity. The training wheels are off').

Worked example · from the session

Scrape Creators wired in the same way: docs URL → 'install this for me' → Codex reads the quick-start and configures the env key; when Hermes later reports the key missing, Codex fixes it while he keeps teaching.

Why it matters

It's the mature version of tool adoption: one trusted harness that reads documentation, so every other tool costs a paste and a sentence.

People get this wrong

Each tool must be learned through its own interface and docs.

One harness that reads docs is the interface to everything else — the marginal tool costs a URL and a sentence, plus the tokens.

It's exactly like asking, can my gardener cut the grass? The gardener can cut the grass if you give him the tool.
I don't want to show you a trailer of MasterChef. I will bring everything together and cook the food in front of you.
For your projects

This is also how THIS project works — Cowork as the harness that installs, debugs, and switches between everything else. The $3-5 budget rule is worth adding to your standing tool-trial habits.

This idea elsewherecomplementsHarshith talks to his Hermes FROM Codex in the Catalyst track — same pattern, learner-facingcomplementsthe sprint economics of choosing which harness burns whose tokens
Go deeper

In one line: Agent-as-harness = use a supervised CLI agent (Codex/Claude Code) as the installer, configurator, and debugger for other agents and tools (Hermes, Ollama, Scrape Creators): paste docs/repo URLs, delegate setup, debug from outside the failing system, switch surfaces freely; guard with small prepaid tool budgets and early steering.

Non-technical is a valid input: 'I am non-technical, set this up on my behalf' is a legitimate, effective instruction ()

Local-model sizing rule captured: ~48-64GB RAM for a good 30B model; his 16GB machine gets a small coder and OpenRouter for real work ()

Surface-switching is a skill, not a defeat — the demo's own failure proved the interchangeability thesis ()

Md-file handoffs between Claude Code and Codex named as the token-exhaustion relief valve ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why debug a failing agent from OUTSIDE it?

The failing system's context is contaminated by its failure; the outside harness reads the CLI, the config, and the docs with clean eyes — same reason surgeons don't operate on themselves.

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01The dynamo and the computer: why bolted-on AI produces bills, not productivityDynamo parallel = general-purpose technologies pay off only after process redesign around their new shape (…›

Dynamo parallel = general-purpose technologies pay off only after process redesign around their new shape (decentralized, always-on, feedback-driven), not after in-place substitution; AI adoption that preserves the old workflow produces cost without productivity.

Redesign includes metrics: the YouTube case chose ONE number (average view duration) and subordinated everything to it ()

Persistence through the ugly middle is part of the redesign — the clone-hate comments era preceded the quality inflection ()

Model progress is a step function you position for: Seedance v4 changed the equation overnight for a system already built around clones ()

02Hermes ≡ OpenClaw ≡ (mostly) Claude Code/Codex: the always-on agent categoryAlways-on agent = continuously running, remotely reachable agent process (Hermes/OpenClaw;›

Always-on agent = continuously running, remotely reachable agent process (Hermes/OpenClaw; local or 1-click VPS) distinguished from supervised CLI agents (Claude Code/Codex) by availability and session memory, not by capability; all interchangeable via md-file harnesses, with 'self-improvement' being configuration, not magic.

'The LLMs don't know what they are' doctrine from the Catalyst track reappears as practice: same principles run on any of the four harnesses ()

Folder-scoped like the others: cd into the project, type hermes — 'exactly like how you would say claude code' ()

Available tools out of the box: browser, computer use, skill packs (creative/data-science/DevOps/MCP/media) — permissioned by you ()

Hostinger's 1-click tier is the sandbox recommendation for the nervous: 'a safer sandbox to play with' than your own laptop ()

03Employee on day one: soul.md, user.md, memory.md — onboarding docs for an agentAgent onboarding = role-separated persistent context: soul.md (agent persona + behavioral contract), user.m…›

Agent onboarding = role-separated persistent context: soul.md (agent persona + behavioral contract), user.md (boss profile + reporting preferences), memory.md (project facts, tools, known bugs), skills (per-job capabilities); authored by interview, kept compact, reloaded every session.

The persona spec doubles as an anti-sycophancy contract — pushback and blind-spot surfacing are WRITTEN DUTIES, not hopes ()

Verification habit: ask the agent whether the file exists and what it says — his user.md turned out to live in the memory system's profile layer, found by asking ()

Global vs project scoping is explicit: persona and boss are global; memory.md belongs to the project ()

Cross-agent: the same interview produced files usable from Codex or Hermes interchangeably — the files ARE the employee ()

04The self-improving ghostwriter: feedback → named lesson → patched skillSelf-improving skill = capability spec (voice DNA / design principles) + explicit loop contract (on feedbac…›

Self-improving skill = capability spec (voice DNA / design principles) + explicit loop contract (on feedback: state the lesson, patch the skill, rewrite) + grounding tools (web/X/session search) + release valve against over-constraint; improvement is boss-fed, named, and persisted — never assumed.

Hallucination handled as a curriculum moment: the fabricated example becomes the skill's first permanent rule ()

Session search is the underrated tool: the agent reads its own past sessions to learn how you write and decide ()

Platform constraints are lessons too — the 3,000-char cap patched in after one violation ()

The loop ran even where he didn't write the prompt: the final image's generation prompt was composed entirely by the system — 'that is the loop' ()

05Running the agent through an agent: Codex as installer, debugger, and escape hatchAgent-as-harness = use a supervised CLI agent (Codex/Claude Code) as the installer, configurator, and debug…›

Agent-as-harness = use a supervised CLI agent (Codex/Claude Code) as the installer, configurator, and debugger for other agents and tools (Hermes, Ollama, Scrape Creators): paste docs/repo URLs, delegate setup, debug from outside the failing system, switch surfaces freely; guard with small prepaid tool budgets and early steering.

Non-technical is a valid input: 'I am non-technical, set this up on my behalf' is a legitimate, effective instruction ()

Local-model sizing rule captured: ~48-64GB RAM for a good 30B model; his 16GB machine gets a small coder and OpenRouter for real work ()

Surface-switching is a skill, not a defeat — the demo's own failure proved the interchangeability thesis ()

Md-file handoffs between Claude Code and Codex named as the token-exhaustion relief valve ()

Tools referenced

ToolCoverageMomentContext
Hermes agent (Nous Research)demonstratedCLI-first; soul/user/memory setup, skills, session search, self-improvement loop; Telegram-connectable; MIT open source
Codex (OpenAI)demonstratedThe harness: installs Hermes, wires Ollama, researches file conventions, fixes keys, debugs, and finishes the image skill when Hermes stalls
OllamademonstratedHooked to Hermes via Codex; 16GB RAM sizing honesty; 'my system will die if I use local models' — teaching demo only
OpenRouterdemonstratedHermes' provider (Opus 4.6 for price); also the image-generation route; credits checked mid-debug
Scrape CreatorsdemonstratedSingle API over TikTok/Instagram/YouTube/LinkedIn/X/Reddit+; free 100 credits, ~$47 paid; installed by pasting its docs into the harness
FirecrawldemonstratedPre-scraped outskill.com brand page reused to feed the color scheme when screenshots overflowed the image API
Parallel (search MCP)demonstratedThe web-research tool Codex is steered to use ('use parallel search to find the answer')
Wispr FlowdemonstratedThe dictation layer for the soul.md/user.md interviews
Claude CodeexplainedDeclared interchangeable throughout; the supervised lane alongside Codex; md-file handoffs relieve token exhaustion
Hostinger / SiteGroundexplained1-click Hermes/OpenClaw installs — the always-on hosting lane and 'safer sandbox'
Visualize Value (Jack Butcher)explainedDesign lineage for the explainer-image skill; his found md file was a writing system, so the principles article was used instead
HeyGenmentionedWhere Seedance v4 lives — the clone-quality step-change in the Viber YouTube case
SeedancementionedThe v4 model that inflected the AI-clone channel's quality ('quite costly, but worth it')
Perplexitymentioned'If your question starts with does Hermes—, ask Perplexity' — the training-wheels boundary
ApifymentionedThe comparison point Scrape Creators replaces ('I'd have to remember which scraper is which')

Action items

Resources mentioned

Resources
  • docHermes agent GitHub (Nous Research)
  • docThe Dynamo and the Computer (Paul David, 1990)
  • docScrape Creators
  • docHostinger 1-click agent installs
  • docVisualize Value principles article
  • docCertificates form + recordings
  • docNo prompt handouts (by design)

Extraction notes

This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

Transcript corrections applied

The transcript saysThe trainer actually means
Duleep / Dilip / the leap / Dileep / Philip / TullyDileep (KVSS Dileep)
Hermos / Hermit / Hermitz / Harmis / Harmeet / her miss / Thomas / phonesHermes (agent)
NOS Research / Nose ResearchNous Research
OpenClaw / open clock / open clogOpenClaw
codecs / Corex / Porex / prod code / plot code / clod codeCodex / Claude Code
sol dot m d / sold dot m d / soul dot m bsoul.md
Ola / Ola Ma / Olama / OLA MAOllama
gamma 3 / gamma 4Gemma (Google local model, as-heard version)
Viber / Vibho / Vaibow / Vibas / VibhavVaibhav (Outskill founder; the AI-clone YouTube channel)
GajinGhajini (the Indian Memento)
Zara Zhangas-heard attribution for the dynamo-paper tweet
Jack ButcherJack Butcher (Visualize Value)
psycho fencing / therapeutic flosssycophancy / therapeutic fluff
agent released the v 4 model with the seed danceSeedance v4 model release (as-heard; available in HeyGen)
Boris... I don't write prompts, I do loopsas-heard attribution (Boris — likely Boris Cherny of Claude Code) for the loops line
Kartik / Philip (closing host)Kartik (sprint host)
Opus 4.6 / 4.8as-heard Claude Opus version numbers at recording
brass taxbrass tacks

True on recording day — verify before relying