Skills

Skills are reusable instruction modules. They let you capture how an agent should perform a recurring kind of work without repeating the same guidance in every conversation.

What a skill contains

• A name and slug that make the skill easy to identify.
• Markdown instructions that describe the expected behavior.
• A scope: global or workspace.
• An invocation mode: automatic, manual, or tool-initiated.
• A priority that decides which instruction wins when multiple skills overlap.

For the shipped default skill catalog, see the Skill Reference.

Global vs workspace skills

Global skills

Global skills are shared defaults that can appear across workspaces. They are the right place for broad house style, common output patterns, and organization-wide review rules.

Workspace skills

Workspace skills belong to one workspace. Use them for team-specific procedures, vocabulary, or output formats that should not affect anyone else.

If a workspace needs to customize a global skill, create a workspace skill with the same slug. The workspace version takes priority for agents in that workspace.

How skills are applied to agents

• Skills are assigned from the agent editor.
• Global skills can start enabled by default for an agent.
• Workspace skills can be turned on or off per agent.
• Each assignment has a priority, so the agent can resolve overlapping instruction sets predictably.
• The context preview should reflect the actual effective skill set, not just the raw list.

Preloaded Skills

Preloaded skills are included in the initial agent context instead of waiting for Dawn to load them later during the run. Use preload for compact, high-value instructions that should shape the agent from the first response, such as a core review format or a workspace-wide operating convention.

Preload can be configured at two levels:

• Workspace defaults: skills that should be preloaded for agents unless an agent opts out.
• Agent-specific preloads: skills that are especially important for one agent profile.

Keep the preloaded set small. Large skill bodies increase the initial context for every run that uses them. Skills that are only sometimes relevant should usually remain normal assigned skills so Dawn can load them when needed.

If a workspace default is inherited by an agent, turning the skill off removes it for that agent. Turning the skill back on should restore the inherited default behavior unless you explicitly change the agent-specific preload setting.

Invocation modes

Automatic

Always active for the agents that have the skill enabled. Use this for default formatting, response structure, and other consistently desired behavior.

Manual

Only activates when the user explicitly asks for it in the message using the skill slug.

#skill:release-notes
Summarise the last 7 days of shipped changes for customers.

Tool-initiated

Activates from the agent’s own tool path rather than behaving like a hidden always-on default. Use this when the skill is only relevant during a specific tool-driven workflow.

Writing good skills

• Keep one skill focused on one concern.
• Write instructions the agent can actually follow and the team can actually verify.
• Include output expectations, not just abstract intent.
• Copy a global skill into a workspace when local customization matters instead of editing around it conversationally every time.
• Review priorities when multiple skills shape the same output.

Agent-Assisted Skill Authoring

Dawn agents can help draft and maintain workspace skills when the workspace allows catalog authoring tools. This is useful when you want the agent to turn a repeated instruction into a reusable asset instead of leaving it buried in one thread.

Typical requests include:

Create a workspace skill for how this team writes PR review findings.

Update the release-notes skill so it always separates customer-facing changes from internal maintenance.

Copy this global support-triage skill into the workspace and adapt it for our escalation policy.

Agent-authored skill changes are still workspace asset changes. Treat them like edits made through the UI:

• review the proposed name, slug, scope, and markdown body;
• keep each skill focused on one behavior;
• test the skill with an assigned agent before relying on it;
• use revisions or audit history when investigating why a skill changed.

Stale Edit Protection

When Dawn updates an existing skill, it should work from the current version rather than overwriting a newer edit accidentally. If another user or agent changed the skill after it was read, Dawn may ask to reload or reapply the change.

This protects shared workspace behavior from stale writes. If an update is rejected, open the latest skill version, review what changed, and ask Dawn to apply the edit again against the current content.

Validation checklist

• Use a test request with an obvious expected format and confirm the output matches the effective skill list shown in the agent editor.
• Global and workspace scope are visible and understandable.
• Workspace overrides with the same slug behave as expected.
• Manual skills only activate when explicitly requested.
• Tool-initiated skills do not silently behave like a hidden global default.

List active skills for this agent and show effective priority order.

Identify conflicting skill instructions and suggest fixes.

Draft a workspace skill for this repeated instruction and explain what I should review before saving it.