Concept
Skills: where the know-how lives
Saved, reusable procedures (SKILL.md files) the agent loads for a particular job; posters treat them as where a business's know-how ends up.
Across six posts, skills are described as the way a Hermes agent turns one-off work into repeatable procedure. Berthelot defines them as markdown files with YAML frontmatter acting as procedural memory, loaded in tiers to keep token cost low, and maintained by a curator that archives unused ones. Business posters use them more practically: the Halfbit author saves a competitor analysis, event research and a strategy advisor as skills so a new team member can load the company’s thinking; Onyeije built 94 narrow clinical and content skills and says focused skills beat vague general intelligence; the GTM poster keeps one skill per job in a GitHub repo; Bara writes down each working routine as a skill. The shared pattern is: do the task once, then save it. Evidence is entirely self-reported, and the one quantified business case (five skills at Halfbit) is explicitly early-stage. Berthelot’s claims about uniqueness among agents and about GEPA optimisation come from a guide, not from observed results.
Key points
- Save a working routine or analysis as a named skill so it can be rerun — 4 posts. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L77) (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L54) (@coldemailchris · Our team built an AI agent that builds our… ¶ L34) +1 more
- Narrow, task-specific skills beat a general assistant — 2 posts. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L68) (@coldemailchris · Our team built an AI agent that builds our… ¶ L36)
- Skills are the procedure layer, separate from memory (facts) and soul (identity) — 1 post. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L70)
- Skills can serve as shared team playbooks that onboard new people or drive jobs — 2 posts. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L137) (@coldemailchris · Our team built an AI agent that builds our… ¶ L35)
- Skill libraries compound: later skills get cheaper to build — 2 posts. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L170) (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L158)
- Skills need upkeep: tiered loading, a curator and offline optimisation — 1 post. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L101)
Where posters agree
- Posters who describe skills agree they hold repeatable procedures that the agent reloads, rather than one-off prompts. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L100) (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L77) (@coldemailchris · Our team built an AI agent that builds our… ¶ L35) +1 more
Frameworks
- Skill progressive disclosure — Three levels that keep token cost low when the agent loads skills. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L101)
- Level 0: names and descriptions only (~3k tokens for the full catalog)
- Level 1: full skill content when needed
- Level 2: specific reference files within a skill
- Skill self-improvement loop — The agent meets a problem, solves it by trial and error, saves the approach as a SKILL.md file and loads it next time. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L116)
- Triggers: completes a complex task; hits errors and finds a solution; user corrects its approach; discovers a non-trivial workflow
- Curator: automatic transitions (stale after 30 days, archived after 90) and LLM review (keep, patch, consolidate, archive)
- GEPA (genetic-pareto prompt evolution) — An offline optimisation pipeline in a companion repo that reads execution traces to see why things failed, then proposes improvements through evolutionary search, scored by LLM-as-judge with rubrics and never committed directly. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L135)
Tactics
- Package a one-off research task as a named skill so it can be rerun with one command. — The competitor analysis was saved as a competition-analysis skill, to be rerun in six months without rebuilding the context. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L77)
- Use a skill to monitor industry event calendars against budget, location and fit criteria. — The event-research skill searches a dozen-plus sources, filters events and recommends, for example a single trip covering two events. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L114)
- Save every process, analysis and decision as a reusable skill so new team members can load the company’s strategy. — A new team member loads halfbit-advisor. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L137)
- Store a skill per GTM job in a GitHub repo the agent reads first — Hermes has repo access; before a job it pulls the matching skill so it works with every written playbook loaded. (@coldemailchris · Our team built an AI agent that builds our… ¶ L35)
- Pin critical skills to protect them from the curator. — The curator never auto-deletes; anything can be recovered from the archive and critical skills can be pinned. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L128)
- Turn a rough voice memo into a set of clinical documents with a signout-to-APSO skill. — The skill outputs a physician-facing sign-out, patient-friendly explanation, referral note, structured outpatient workup plan, list of missing data, care coordination plan and a draft message to the referring provider. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L110)
- Use GitHub skills to ask the housekeeping questions a solo builder forgets. — The assistant asks whether there is a README, documented environment variables, a deployment checklist, a security note, and whether the feature ties to a real clinical workflow. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L162)
Stories & examples
- Enterprise consulting methodology: Considering consulting for larger companies, the author used the agent to build surveys for several groups, a data-gathering process, a hypothesis matrix and a 4-day workshop format, plus traps to avoid. Lesson: The skill holds a ready-made process the author says could be sold to a company with 500 developers. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L119)
- Skill library compounds: In week one the author spent as much time building skills as using them; by week two the ratio flipped and early skills ran daily without modification. Lesson: New skills become cheaper to build because patterns already exist. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L170)
- Lead research for 24 leads: Sales began as an Excel list with no research; the agent researched 24 leads and created a YouTrack ticket for each. Lesson: The lead-research-brief skill turns a company name into an outreach brief. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L83)
- List build from one Slack message: The post links a video of Hermes building a contact list from a single Slack message (the video is not in the text). Lesson: A single request can drive a list build when the playbooks are loaded as skills. (@coldemailchris · Our team built an AI agent that builds our… ¶ L37)
Numbers
- Specialised skills in the build: 94 skills (self-reported) (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L65)
- Skill catalog size at Level 0: 3000 tokens (self-reported, approx.) (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L103)
- Unused skill thresholds: 30 days stale, 90 days archived days (self-reported) (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L124)
- Business skills so far: 5 skills (self-reported) (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L154)
- Time for a new team member to match the author’s strategy knowledge: 5 minutes (self-reported) (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L138)
Caveats & warnings
- The author says the company is at an early stage with only 5 business skills. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L153)
- No other open-source agent ships runtime skill learning, persistent multi-layer memory and an optional weight-training pipeline together. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L31)
- Getting the right voice from content skills needed more iteration than anything else, even though documentation is more technically complex. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L176)
How much to trust this
- Evidence: 6 posts, 6 independent speakers (repeat posters counted once).