An Agent That Stops Recomputing: Building AI Skills That Actually Work on Hermes

The narrative opens on a single frame: an agent that truly works needs three pillars — a solid structure, a server to keep it alive, and skills that automate the job. The pain point is immediate: writing those files by hand is slow and fussy, which is where most people stall. From that gap comes the promise of a free generator, with the claim that anyone following to the end will have a live skill running for free on their own agent.

Why an agent without skills drifts on every rerun

Without skills an agent derives its approach from scratch on each request, and the cost shows in repeatability. Run the same prompt twice and the first output returns with a source beside every line and a decision matrix at the bottom; the second comes back shorter, citation-free and with different facts. The method used in the first run dies when the session ends, a fresh method is invented next time, and there is no reference to make the two line up. Swapping in a larger model does not repair this, because the source of variance was never model size but the absence of a reference.

That weakness is tolerable for one-off questions where you can read and re-ask, but it collapses once you want the same job every week without supervision. A skill fixes that by becoming the reference the agent can open and follow. Physically it is a folder on the agent's machine, one per skill, holding a SKILL.md file that starts with a short applicability description and then the steps; optional code files, reference folders and templates sit alongside for jobs that need more than prose, while a simple skill is often that single file.

An agent reads none of it until a job matches the description, so a library of skills sits inert at near-zero cost until called, which makes it sensible to add them one by one without interference or a ceiling to chase. The demo runs on Hermes, an open-source agent from Nous Research cited in the video at about 236,000 GitHub stars — our checks show 240k+ and a TechCrunch report of talks to raise at a $1.5 billion valuation — MIT-licensed so what you build stays runnable even if the vendor disappears, with the added note that Hermes can also draft skills itself from work it already completed well.

The two backbone files: identity and skill

Before handing over real work, the video says two backbone files must be in place, the first being identity. On Hermes that is soul.md, the primary identity injected as the first block of the system prompt on every run. The install walks to /opt/data/soul.md in the demo, while the official docs list the canonical home as ~/.hermes/SOUL.md (or $HERMES_HOME/SOUL.md); Hermes seeds a starter file on first run and falls back to a built-in default if the file is missing or empty. Whatever sits in that file shapes tone, autonomy and refusal behavior; procedure tells the agent how to do the job, identity tells it how to behave all the while, so two agents on the same skill but different identities hand back different results.

Writing pages of rules from a blank file would stop most people, so the video introduces Ordain at ordain.host: enter an email, receive a one-time code and start building. The site has two builders, soul first because a skill needs a host agent. At the top you pick the platform — Hermes versus OpenClaw in this case — then a name, then check boxes for what the agent handles; the narrator ticks messages and email, research, summaries, reminders and scheduling while leaving content creation, customer replies, coding, help and general assistant off, noting that extra ticks dilute focus. Personality offers professional, friendly, direct, warm and playful — direct is chosen — and autonomy spans from ask-before-everything to handle-small-stuff-ask-on-big to just-get-on-with-it-and-report-after, with the middle setting preferred. The free-text box at the bottom is presented as the place worth the most time: asking for British English, answer-first with reasoning below, no list padding to chase a round number and an explicit statement of uncertainty when unsure — vague prompts make vague agents, generic virtues such as helpful teach nothing new, and hitting generate fills the right pane with a page of instructions in the agent's own voice ready to download as txt.

Why a server is not optional and how it is provisioned

Even with a polished identity, there is no agent without a machine that stays on; when the box sleeps the agent vanishes and that matters doubly for skills whose value peaks while you sleep — the research skill walks a stack of pages alone at 06:00, impossible on a closed laptop. The host in use is Hostinger's KVM2 with two virtual cores, 8 GB RAM and 100 GB NVMe plus a one-click Docker deploy for Hermes; the link drops you to the deploy page with Parker's coupon already applied and a 12-month term suggested. After payment the project appears under Docker Manager alongside Traefik for networking, the name ends with four random characters, the top-right web terminal opens and two commands are pasted — one enters the container, the other runs hermes setup with full setup, OpenRouter and a ~$25 cap. Telegram is then connected, the soul txt and Ordain copy prompt pasted back-to-back, the agent writes its own soul.md to the correct folder and a /reset starts the new session where the file is finally read, a step people skip so changes never apply inside the session they were made.

The one box that makes or breaks a skill

Whether a skill works comes down to one box in the second builder: what should it do. Type a thin line like research topics and give me a summary and you get a file that roughly says that, leaving the agent to keep making the same choices as before, now with a file that mirrors them. The version that works names the sources to start from and the order to walk them, states what the finished brief must contain and spells out what to do when sources conflict or when nothing is found; that final clause stops the agent from filling a blank with a plausible number. A requirements box underneath locks cross-run guarantees — in the example, every claim carries a link and nothing not opened is included.

Hitting generate again produces a proper skill shaped to the chosen platform. The top is front matter with name, a one-line applicability description and tags; that line does more work than it looks because it is the only part the agent reads when deciding whether to open the rest, so a vague description means the skill is never opened even on the job it was built for. The body then runs through when to use it, the step-by-step procedure, pitfalls and how to verify completeness at the end. The pitfalls section captures mistakes already observed in the wild and suppresses them going forward, while the verification clause — often skipped — is why a well-written skill can still return badly formatted work. Delivery is two-way: upload the txt in Telegram and paste the copy prompt so the agent writes the folder under hermes/skills itself, or drop the downloaded zip directly and the skill goes live instantly; being a text bundle, edits are just open, change a line and replace.

Trust is earned on a known-answer task first. In the demo the agent completed the job correctly but formatted wherever it wanted, ignoring the file's template; the steps were too loose, not the model wrong, and adding a line that mandates the exact printed format fixed the second run — that is how the pitfalls section grows. The same research job from the opening is then given word-for-word; the agent matches the description, opens the procedure and prioritizes official docs and change logs before any forum thread. The returned brief leads with a one-line summary, then every change from the last 90 days with a link beside each figure, disagreements written with both sides kept and nothing dropped, and a closing list of what could not be confirmed, each entry noting where it looked rather than inventing.

Run the same job again and the headings and order recur, sources and sections identical, only findings differ; nothing was touched between runs and that second run is the one to watch, because the first only proves the job can be done once — which it could before any skill. The final step is scheduling: ask it to run every weekday at 06:00 and message the brief on Telegram and it wires that itself. The video's starter advice is to build four skills before anything else — research for fastest payback on repetition, inbox that drafts in your voice for you to edit rather than compose, content that returns a script in your weekly structure instead of a model's assumed one, and outreach that finds candidates and drafts the opener you never get around to. The heuristic for choosing more is practical: anything you have done the same way more than three times encodes well as a skill, anything where the path diverges on what you find in the first ten minutes does not, and forcing it makes the agent worse; you can stack many because only matching descriptions are opened, so the library can grow without slowing runs, quietly moving your working style from memory into files until the agent is no longer a tab you visit but a routine that ran while you slept, which is why the demo's hands-free run and the closing Hostinger KVM2 pitch with code Parker are presented as a single loop.