Skills and instructions

Custom instructions are shared free-form guidance. Skills are named packages with a description and content that Tasuku places into the provider's expected skill directory.

Provider packaging

For Codex, skills are exposed through the runtime's .agents/skills convention. For Claude, Tasuku uses the provider-specific .claude/skills location. The logical skill stored in Tasuku remains provider-neutral.

Each skill includes:

• a stable name;
• a short description used for discovery;
• instruction content applied when the agent selects or receives the skill.
• an optional Disable model invocation policy.

When Disable model invocation is enabled, Tasuku marks the skill as user-only for providers that support that policy. The model cannot select it automatically; workflow or user instructions must invoke it. Providers without an equivalent control still receive the skill package, so do not treat this flag as a security boundary.

Bundled templates

The Skills tab includes four templates that can be added with the plus button and then saved with the rest of the configuration:

• Karpathy Guidelines;
• Thermo-Nuclear Code Quality Review;
• Code Review and Quality;
• Security and Hardening.

Tasuku stores a copy of the selected template, including its source attribution and invocation policy. Adding a template does not modify the upstream source later, and a template with the same skill name cannot be added twice. Source links and MIT notices are available directly from the gallery.

Scope

Instance skills and instructions are defaults. Repositories may override the entire skills category or instructions category independently. Agent behavior is part of the instructions category. A repository override is not merged item-by-item with the instance category.

Guidance quality

Keep instructions specific to observable repository practices. Do not put credentials in skill content. Use Secrets and ${secrets.NAME} references for supported configuration fields.

Learnings are separate from manually authored skills. When a trusted human explicitly states a durable preference in a Tasuku-owned GitHub review thread, Tasuku records and enables it immediately with conversation provenance. Administrators can disable a learning from the dashboard when it should no longer affect future runs.