Skills System

Skills are reusable instruction packages that teach Anyy a focused workflow. A Skill can contain a SKILL.md entry point, linked supporting files, and a command exposed in compatible chat surfaces.

Skills belong to the selected Profile. Each Profile has its own catalog and enabled state.

Before You Install

Read the source, especially its commands, external services, credential expectations, downloads, and linked files.

Treat Skills as executable guidance

Installing a Skill does not itself grant new operating-system permissions, but using it can direct Anyy to call enabled tools. Built-in scanning can reject known-dangerous patterns; it cannot prove that arbitrary instructions are safe.

Understand Skill Sources

Dashboard groups Skills by origin:

Source	Meaning
System	Bundled with Anyy; visible but not user-toggleable or removable
User Installed	Installed into the Profile's managed Skills root
Agent Created	Managed provenance label for an existing Skill record created by an agent-capable workflow

Inspect a Skill

Open Dashboard → Skills, then select an item to review:

• ID, name, description, and scope;
• source and managed state;
• enabled or disabled state, including a displayed reason when unavailable;
• rendered SKILL.md content;
• file path; and
• linked supporting files.

From the CLI:

anyy skills list
anyy skills view SKILL_NAME
anyy skills view SKILL_NAME FILE_PATH

Enable or Disable

Use the Dashboard toggle for a non-system Skill. Disabling removes it from normal resolution without deleting its files. Re-enable it when the workflow should be available again.

If a running surface still shows an old command catalog, reload Skills and start a new session:

anyy skills reload

Install from Dashboard

Select Install Skill and choose a source.

GitHub directory

Enter the Skill ID and a source in this form:

github:owner/repository/path@revision

The optional path selects a directory inside the repository. Pin a revision when you need a reproducible install.

Local zip

Upload a .zip containing the Skill. Dashboard accepts up to 10 MiB, stages and validates the archive, and rejects path traversal, symlinks, and entries outside its staging root.

Both Dashboard install paths submit a high-risk ChangePlan. Files are not committed until you approve it.

Install from the CLI

The CLI also accepts an accessible local directory or an HTTPS URL that points directly to SKILL.md:

anyy skills install github:OWNER/REPO/PATH@REF
anyy skills install --all github:OWNER/REPO

A single-file HTTPS install needs an explicit Skill name when the source cannot provide one. Remote URLs must use HTTPS and must not contain embedded credentials.

Useful options:

• --force replaces an existing managed Skill;
• --now reloads the resident Gateway after the mutation; and
• --json returns machine-readable output.

Use Skills in Chat

Ask for a workflow by name or describe the task it covers:

Use the release-notes Skill to summarize the changes in this Workspace.

Type /skills in the TUI or / in Dashboard Chat to browse the live catalog. The exact commands depend on the Profile, surface, installed Skills, and enabled state.

More than one Skill can participate when a task genuinely needs them. Loading unrelated Skills adds instruction volume and can create conflicting guidance.

Skill Management by Anyy

When the Profile has a managed Skills root, ordinary sessions also expose the audited skill_manage tool. Anyy can use it to install, create, patch, edit, delete, roll back, reload, or maintain supporting files when the task explicitly calls for Skill management.

Unlike Dashboard installation, this model tool does not create a separate ChangePlan or approval prompt. It is classified as a high-risk configuration change and every execution is audited. Review the resulting Skill before relying on it, and use rollback or removal when the change is not wanted.

System Skills remain immutable. Agent-managed changes apply to the selected Profile's managed Skill library.

Automatic Self-Improvement Review

Anyy keeps a per-session counter of eligible foreground model-loop iterations. When that counter reaches 15 and skill_manage was available to the originating session, Anyy can start a separate Self-improvement review background session.

An iteration is one pass through the model loop, not one user message. A single tool-using turn may contain several iterations and reach the threshold sooner than 15 chat turns. Foreground executions ending as completed or max_iterations count; Cron runs, delegated child work, retries, other failed turns, and the review itself do not. If the foreground run used skill_manage, the counter restarts after that management action and counts only later iterations.

The review receives a bounded transcript from the triggering execution and uses the same selected provider and model. Its tool catalog is narrowed to:

• skills_list to inspect the catalog;
• skill_view to read relevant instructions; and
• skill_manage to update Profile-managed Skills.

It cannot use Memory, terminal, messaging, provider administration, MCP, or delegation tools. The review looks for reusable workflow lessons, corrections, and preferences that belong in a class-level Skill. It may update an existing Skill, create a managed Skill, add a supporting file, or conclude that there is nothing to save.

The review can change Skills without an approval prompt

The same audited skill_manage policy applies to the background review. Inspect Dashboard → Chat → Background, Dashboard → Skills, and Audit after a change. A successful mutation also produces a Self-improvement review notification; a no-change review does not announce a Skill update.

The separate review is a real provider request and can add tokens, latency, and cost. The current version does not expose a public setting to change or disable the 15-iteration interval; the review is skipped when managed Skill editing is unavailable to the session.

This mechanism is not personal Memory. A stable fact about the user belongs in USER.md; a reusable method for a class of tasks belongs in a Skill. See Persistent Memory.

Roll Back or Remove

Roll back to the previous managed revision or select a recorded revision:

anyy skills rollback SKILL_NAME
anyy skills rollback --revision REVISION_ID SKILL_NAME

Remove a managed Skill:

anyy skills uninstall SKILL_NAME

Dashboard can submit removal for a managed Skill. System Skills cannot be removed. Add --now to a CLI rollback or uninstall when the Gateway should reload immediately.

Without an explicit revision, rollback selects the previous managed revision. Offline mutation is allowed only when the Profile is not owned by a resident Anyy runtime.

Tips

• Inspect a Skill again after an update; its instructions may have changed.
• Pin a GitHub revision when reproducibility matters.
• Keep durable personal preferences in Memory, not in a Skill.
• After an automatic review notification, inspect the exact Skill change and roll back if it generalized the lesson incorrectly.
• Use Skills for repeatable procedure; use MCP for an external tool or service connection.
• A Skill does not grant tools or permission by itself; see Tools & Toolsets.

Troubleshooting

Problem	Fix
Install or removal is pending	Review the ChangePlan under Approvals
Skill does not appear in Chat	Enable it, run anyy skills reload, and start a new session
Zip is rejected	Remove symlinks and path escapes, and keep it at or below 10 MiB
Rollback or uninstall is blocked	Use the resident workflow, or stop Anyy before an offline mutation
System Skill cannot be toggled	System Skills are managed by the Anyy release
A Skill changed unexpectedly	Check the Background review session and Audit, then use anyy skills rollback SKILL_NAME
No review appears after 15 chat messages	The trigger counts eligible model-loop iterations, not messages, and requires skill_manage in the session catalog