Summary

Hermes learns procedures by saving and revising instructions. Review what it saves and compare results before adopting a change.

What you'll learn

  • Separate a corrected output from a saved rule.
  • Review procedures before sharing them.
  • Retest when work or tools change.

What does Hermes save as a skill?

As of October 2026, Hermes can turn a reusable solution into a skill: written instructions for a task, with optional supporting files. The required file is called SKILL.md. It explains when the skill is useful and how to perform the work. Scripts, reference documents and templates can accompany it; include only the supporting resources that the procedure needs.[1]

The useful distinction is between completing a task and preserving an approach. An ordinary successful reply may not need a new skill. Solving a nontrivial problem, discovering a repeatable method or incorporating an important correction can make saving or updating one worthwhile.

For a recruiting workflow, that might mean a reusable checklist for checking an application against agreed criteria. The skill could explain how to handle a missing attachment and where to escalate an unclear case. It does not need to store every applicant’s personal information to preserve that method.

Skill directory with a required SKILL.md instruction file and optional scripts, references and templates
A skill is a reusable instruction package. Its supporting resources depend on the task.

How does Hermes decide what to learn?

As of October 2026, Hermes can create or revise skills when a useful, reusable lesson emerges. Background review can also propose or stage changes after a session.[1] The agent has a skill-management tool for this work. Saving is selective: a task finishing does not guarantee a new skill, and an edited output does not guarantee that the edit becomes a persistent instruction.[1]

That matters to a business reviewer. If you correct a client report, distinguish the correction to this report from a change you want applied to future reports. Ask for the reusable instruction to be updated, then inspect what was actually saved. Otherwise the next report may repeat the same mistake.

Saving a skill changes the instructions available to the agent. It does not retrain its underlying AI or prove that the new instructions are better. Confirm the saved change and test it on representative work.

Hermes also offers an optional human approval step for skill writes, off by default.[1] This controls creation and revision of skill files; it is not a universal approval gate for every email, payment or change in another application. Those actions need their own access restrictions and review rules.

Do more tasks guarantee better performance?

More tasks can expose useful variations, but volume alone does not make the agent more accurate. A repeatedly reused mistake can stay a mistake. Improvement requires a relevant lesson to be saved correctly, applied to the right task and checked against current expectations.

Use milestones based on evidence rather than a promised month-by-month curve:

MilestoneEvidence to look forDecision
First reusable approachThe saved instructions describe the actual task and its limitsReview before reuse
First revisionThe changed instruction addresses an observed failureTest both the original case and similar cases
Repeated useOutputs meet agreed requirements without introducing a new error patternKeep the skill; retain review for consequential actions
Changed workflowThe old instructions no longer describe the tools or standardRevise or retire the skill

Track correction and escalation rates on a consistent sample of work. Record the skill version and any model change. An improvement in one task category is evidence about that category, rather than proof that the whole agent has become reliable.

Evaluation sequence: establish a baseline, revise a skill, compare outputs, and keep or revise the change
Progress is a tested result, not a guaranteed curve tied to time or task volume.

Where are Hermes skills stored?

As of October 2026, skills normally live in the Hermes profile’s local skill directory. The usual path is ~/.hermes/skills/; other installation methods and profiles can use a different data location. You can also configure external or project-specific skill directories. Keep this storage in your backup and access-control plan.[1]

Agent Skills is the open format used for these instruction packages. Its website describes the format. It is not a service that automatically receives your task history or a mandatory registry account for your deployment.[2]

You can deliberately copy or distribute a skill to another Hermes installation or a compatible tool. Before doing so, inspect it for client information, internal references and assumptions about available tools. A transferable file format does not make every instruction safe or usable in another environment.

For example, an application-checking skill can be shared between two offices after removing personal examples. Each office still needs the appropriate system access, local criteria and test cases. Transfer the relevant rules explicitly and test how the receiving agent applies them.

A reviewed skill directory is deliberately copied from one Hermes instance to another compatible installation
Sharing transfers selected files. Review their contents and test them in the receiving environment.

What affects Hermes skill quality?

Review instructions for accuracy, corrections for reusable meaning and reuse for the right scope. These three proposed checks help you decide what to keep or revise.

Accurate instructions. Use representative examples and make the limits explicit. A reporting skill should identify the figures to include, their sources and what to do when a source is missing. It should not infer a missing figure simply to complete the report.

Useful corrections. Explain the requirement behind an edit. “Use the client’s agreed reporting period” is reusable guidance; silently replacing one date may only fix the current output. Check that the saved instruction reflects the intended general rule.

Controlled reuse. Apply a skill only where its assumptions still hold. A checklist for one client’s renewal process may be inappropriate for another client. Decide which parts are shared and which belong in separate instructions.

A solo founder and a team both need a way to inspect changes, identify repeated errors and decide whether a skill should continue to run. Keep review proportional to the consequences of a bad output.

How do you maintain skills when work changes?

A skill file can remain intact while the workflow around it changes. A CRM field may be renamed, a client may change its reporting standard or the chosen AI may behave differently after an update. A rise in edits is a reason to investigate, not a diagnosis that only the instructions can be at fault.

Observed problemWhat to inspectNext action
Wrong output formatSaved instructions and current output standardRevise the relevant instruction and test
Failures after a tool changeField names, permissions and connection behaviorRepair the connection and replay representative cases
Repeated exceptionEscalation rule and coverage of unusual inputsAdd the missing rule without broadening access unnecessarily
Failures after a model changeOutputs before and after the changeCompare with the same test set before accepting the update

Preserve a working version before changing a consequential skill. Test a revision against ordinary inputs and the case that prompted the change. Keep human review where an incorrect result could harm a client, spend money or change an important record.

For the deployment decisions behind this process, see the Hermes setup guide. For its broader capabilities and limitations, see what Hermes is.

Frequently asked questions

How does Hermes get better at tasks over time? Hermes can save a reusable approach as a skill and revise it when a useful lesson or correction warrants a change. It does not create a new skill after every completed task. Improvement depends on what is saved, whether it is correct, and whether the skill still matches the workflow.[1]

Does Hermes retrain the underlying AI model as it learns? Saving a skill does not retrain the AI model. It gives the agent written instructions and, optionally, supporting resources to use on later tasks. A separately chosen model update can also change behavior, so compare outputs again when either the skill or the model changes.[1][2]

Can skills be shared between different Hermes instances? Yes. You can deliberately copy or publish a skill directory and install it in another instance. Agent Skills defines the file format; agentskills.io is not the registry where Hermes stores your skills. Review sensitive content and test the receiving instance’s tool access before sharing or using a skill.[1][2]

What happens to Hermes skills if the agent is paused? Saved skills remain in the local profile as long as that storage is retained. Pausing the agent does not remove them. Back up the profile and check that instructions, credentials and connected systems still match the workflow before restarting.[1]

What causes Hermes skill quality to decline after it was working well? An old instruction can stop matching a changed workflow, tool or output standard. A model change can also affect results. Compare the failing outputs, identify which dependency changed, revise the relevant skill or connection, and test the change before using it on consequential work.

Notes

  1. Skills System. Living documentation, October 2026. Accessed October 15, 2026.
  2. Agent Skills overview. Living documentation, October 2026. Accessed October 15, 2026.
  3. Persistent Memory. Living documentation, October 2026. Accessed October 15, 2026.