Create a skill

Execution reference: This page specifies the tool protocol for automation agents. For instructions to follow in Studio, use the user guide.

Create a skill for repeatable instructions. Add a resource when it needs read-only context that can be loaded by name.

Agent-facing flow

  1. Load the Veryfront router skill: veryfront__load_skill({ "skillId": "veryfront" }).
  2. Treat this page as the guide for creating or editing a project skill.
  3. If the request provides no usable task or skill purpose, call veryfront__form_input exactly once with one required brief field. Wait for a usable purpose before creating the skill. Do not choose an unrelated default task.
  4. Create or update the smallest useful skill. Prefer the framework source shape when editing project files. Use the MCP tool call when Studio exposes skill creation tools instead.
  5. If the skill needs stable reference material, add a focused resource instead of pasting long background text into the skill.
  6. After saving, read the saved instructions back with veryfront__read_skill_source for hosted mutations, or reread the saved SKILL.md when editing project source. Verify the name, allowed tools, and instructions against the requested skill. Stop if verification fails.
  7. After successful readback, return the skill path or skill id, any resource path or resource id, and a short note on how to invoke it.

Guardrails

  • Do not load a removed create-skill sub-skill.
  • Do not create UI files, app routes, or unrelated project code for a skill-only request.
  • Do not invent tool names, resource ids, or project paths. Use the tools and conventions available in the current project.
  • Keep one skill focused on one repeatable workflow.
  • Keep resources read-only. Use a tool when the task needs to change state.
  • Do not turn the Mermaid example into the main point unless the user asked for a diagram skill.
  • The bare Create a skill. starter requires the brief form before mutation. It does not authorize selecting a default task.
  • Do not finish with inline-only guidance when file or skill creation tools are available.

Example pattern

This example assumes the user requested an agent-design review skill. Edit the source file or use the available MCP creation tool. Do not substitute this example for a missing brief.

md
---
name: agent-design-review
description: Review AI agent designs for goal fit, tool access, context, handoffs, and verification.
allowed-tools:
  - veryfront__list_files
  - veryfront__get_file
  - veryfront__search_file_text
---

# Agent Design Review

Use this skill when a user asks for a quick review of an AI agent, tool agent, or multi-agent design.

## Steps

1. Read only the agent, skill, resource, or tool files needed for the review.
2. Identify the agent's goal, required context, tools, handoffs, and verification path.
3. Read `agent-design-guide://review` when it is available.
4. Return the main risks, recommended edits, and a small Mermaid map only when the relationships are not obvious.

## Example

```mermaid
flowchart LR
  User --> Orchestrator
  Orchestrator --> ToolAgent
  ToolAgent --> ExternalTool
  Orchestrator --> VerifiedAnswer
```
js
const skill = await veryfront__create_skill({
  "project_reference": "<project>",
  "source_target_kind": "project",
  "name": "agent-design-review",
  "description": "Review AI agent designs for goal fit, tool access, context, handoffs, and verification.",
  "allowed_tools": ["veryfront__list_files", "veryfront__get_file", "veryfront__search_file_text"],
  "instructions": "# Agent Design Review\n\nUse this skill when a user asks for a quick review of an AI agent, tool agent, or multi-agent design.\n\n## Steps\n\n1. Read only the agent, skill, resource, or tool files needed for the review.\n2. Identify the agent's goal, required context, tools, handoffs, and verification path.\n3. Read `agent-design-guide://review` when it is available.\n4. Return the main risks, recommended edits, and a small Mermaid map only when the relationships are not obvious.\n\n## Example\n\n```mermaid\nflowchart LR\n  User --> Orchestrator\n  Orchestrator --> ToolAgent\n  ToolAgent --> ExternalTool\n  Orchestrator --> VerifiedAnswer\n```"
})

const verifiedSkill = await veryfront__read_skill_source({
  project_reference: "<project>", skill_id: skill.id, source_target_kind: "project"
});
if (verifiedSkill.id !== skill.id || verifiedSkill.instructions !== skill.instructions) {
  throw new Error("Saved skill instructions did not match; inspect the source before reporting success.");
}

Both tabs define the same skill. Source allowed-tools maps to MCP allowed_tools. The name and description fields keep their names, and the API writes the frontmatter. The body below it becomes the MCP instructions string verbatim. On veryfront__update_skill, sending instructions replaces the whole body. Omitting it preserves the body.

Resource example

Use a resource when the skill needs reusable context that should be read, not executed.

ts
import { resource } from "veryfront/resource";

const agentDesignGuide = resource({
  pattern: "agent-design-guide://review",
  description: "Review checks for AI agent and tool-agent designs.",
  load: () => {
    return {
      content: [
        "# Agent design review guide",
        "",
        "- Start from the user outcome and the stop condition.",
        "- Keep one agent responsible for one clear job.",
        "- Name required tools, integrations, and resource inputs explicitly.",
        "- Define handoffs, approval points, and verification evidence.",
        "- Use Mermaid only when agent/tool relationships need a map.",
      ].join("\n"),
    };
  },
});

export default agentDesignGuide;
js
// This example targets the main branch. Keep all calls on the same project and branch.
const project = "<project>";
const resource = await veryfront__create_project_resource({
  project_reference: project,
  id: "agent-design-guide",
  title: "Agent design review guide",
  description: "Review checks for AI agent and tool-agent designs.",
  pattern: "agent-design-guide://review"
});
const resourceSource = "import { resource } from \"veryfront/resource\";\n\nconst agentDesignGuide = resource({\n  pattern: \"agent-design-guide://review\",\n  description: \"Review checks for AI agent and tool-agent designs.\",\n  load: () => {\n    return {\n      content: [\n        \"# Agent design review guide\",\n        \"\",\n        \"- Start from the user outcome and the stop condition.\",\n        \"- Keep one agent responsible for one clear job.\",\n        \"- Name required tools, integrations, and resource inputs explicitly.\",\n        \"- Define handoffs, approval points, and verification evidence.\",\n        \"- Use Mermaid only when agent/tool relationships need a map.\",\n      ].join(\"\\n\"),\n    };\n  },\n});\n\nexport default agentDesignGuide;\n";
const before = await veryfront__get_file({ project_reference: project, path: resource.source_path });
if (!before.checksum || !before.version_id) throw new Error("Resource version is unavailable; stop before updating.");
await veryfront__update_file({
  project_reference: project,
  path: resource.source_path,
  content: resourceSource,
  expected_checksum: before.checksum,
  expected_version_id: before.version_id
});
const after = await veryfront__get_file({ project_reference: project, path: resource.source_path });
if (after.content !== resourceSource) throw new Error("Resource content verification failed.");

The creation call returns a scaffold. Write the complete source to its source_path using fresh checksum and version guards, then read it back before reporting completion. If another writer changed the file, inspect that change before retrying.

Output shape

Return the saved skill name, ID or source path, allowed tools, any backing resource ID or path, and one invocation example. Keep the answer brief. An empty resource scaffold is not a completed reference resource.

Next

GoalPage
Understand project skillsSkills
Create a project agentCreate an agent
Understand resourcesResource
Use the resource APIveryfront/resource
Use the skill creation toolveryfront__create_skill
Use the resource creation toolveryfront__create_project_resource