Skip to main content
Trinity
Updated Nov 17, 2026main@87e21ff
Cloud Code Plugins/create-agent

create-agent Plugin

Create new Claude Code agents from an interview (create-agent v2.4.0). The custom wizard asks about your domain and builds a fully configured, Trinity-compatible agent; website scaffolds a site; review, adjust, and clone work on agents you already have.

create-agent 2.0.0 (2026-09-14): the eight pre-built domain wizards (prospector, chief-of-staff, webmaster, recon, receptionist, ghostwriter, kb-agent, doctor) were retired. Describe the same domain to /create-agent:custom and the interview builds it. If you used kb-agent, run /create-agent:custom and then /agent-dev:add-memory for the knowledge graph.

Build an AI Recruiter Agent

Jun 2026

Installation

/plugin install create-agent@abilityai

Available Wizards

WizardCommandWhat it creates
create/create-agent:createDiscovery entry point — shows all wizards and lets you pick interactively
custom/create-agent:customAny agent, from an interview - role, skills, schedules, credentials, Trinity wiring
website/create-agent:websiteSingle Next.js website scaffold (no agent, just a site); also runs headless inside a deployed agent
clone/create-agent:cloneClone an existing agent repository as starting point
review/create-agent:reviewRead-only audit of an existing agent - prioritized findings report, no changes made
adjust/create-agent:adjustApply best-practice improvements to an existing agent

How It Works

Discovery Entry Point

Run /create-agent:create to see the six available wizards and select one interactively.

Wizard Flow

The custom wizard interviews you about:

1.Identity - Name, purpose, personality
2.Tools - Which integrations and capabilities
3.Workflows - Key playbooks and automations
4.Configuration - Environment variables, credentials needed

Output

The wizard creates a complete agent directory:

my-agent/
├── CLAUDE.md              # Agent identity and instructions (carries the playbook-call rule)
├── template.yaml          # Trinity metadata: resources, credentials:, schedules:, plugins:, metrics:
├── .env.example           # Required environment variables
├── .mcp.json.template     # MCP server configuration template (${VAR} in env blocks only)
├── dashboard.yaml         # Metrics dashboard definition
├── .claude/
│   └── skills/            # Initial playbooks
│       ├── onboarding/    # Setup progress tracker
│       └── update-dashboard/
└── .gitignore

Five conventions the wizard bakes in, so the agent is deployable as generated:

•template.yaml declares everything Trinity materializes at creation - the credentials the agent needs (so the guided credential checklist is populated), its recommended schedules:, and the plugins: it depends on (the scaffold declares the plugins it already tells you to install, trinity@abilityai at minimum, so the selection survives a rebuild). See Creating Agents.
•Schedules are one-line playbook calls - a generated schedule message is /daily-briefing, never a prose description of the work; the human-readable intent lives in the schedule's purpose: field. Cron times default to UTC (the container clock) with canonical IANA zone names.
•Business metrics are declared, then recorded - the scaffold adds a metrics: block to template.yaml for the KPIs the agent's domain actually produces, with a cadence that matches the /update-dashboard schedule. The generated /update-dashboard computes the numbers once, writes dashboard.yaml, and records the same numbers as points with the record_metrics MCP tool. Off Trinity, or when the tool refuses, it skips that step silently and the dashboard write still succeeds. A dashboard widget can bind to a declared metric with metric: <name> instead of carrying a hand-typed value. Every metric widget still carries a value, and every status widget a value and a color (placeholders such as "—" and gray are fine), because the platform's compatibility check requires them whether or not the widget is bound. See Dynamic Dashboards.
•.claude/settings.json is not gitignored - it is the agent's project settings and may be committed. Trinity keeps it out of a commit only when it registers platform hook paths or carries credential-bearing keys, so keep secrets in .env.
•Delegation is by playbook call - the generated CLAUDE.md guidelines tell the agent to hand work to other agents only by invoking a named playbook, one line, on any transport. See Playbook calls.

Deploying the Result

The wizard ends by handing off to /trinity:onboard, which deploys from the agent's GitHub repo. With your own GitHub token holding Contents: Read and write on that repo, Trinity gives the agent a working branch it alone writes, with auto-sync on, so its work lands back in git. A read-only token, or the instance-wide token an admin set, gives a pull-only agent that tracks the branch. The create response's git_mode says which you got. See GitHub Sync.

Headless Mode (Inside a Deployed Agent)

/create-agent:customalso runs inside a deployed Trinity agent - a “builder” agent that creates other agents - where there is no interactive question tool. In that mode:

•The brief is the answer. Purpose, name, skills, plugins, credentials, and schedules come from the argument. Decorative choices take their defaults silently.
•Real gaps go out as one ask. Anything the new agent cannot work without and the brief does not say (a data source, an account, a locale) is collected and raised as a single ask_operator question, never guessed.
•One approval before anything leaves the workspace. The skill outlines, schedules, and credentials needed travel in one approval ask. Nothing is pushed or created until a person approves it.
•The agent is created on the builder's own instance. With a usable GitHub token it creates a private repo and calls create_agent from it. Without one it falls back to deploy_local_agent and says plainly that the new agent has no reproducible source and no auto-sync. Either call needs the builder agent to hold the Manage other agents permission, which an instance admin grants in that agent's Settings → Guardrails → Permissions to change itself; without it the create is refused and the skill names the missing permission instead of retrying.
•The approval ask has real endings. It offers explicit options; a person can pick one, answer something else (which approves none of them), cancel, dismiss, or let it expire, and the wizard proceeds only on an approving choice.
•A moot ask is withdrawn. When a new brief answers a question the builder already asked, it ends that pending ask with withdraw_ask; when the facts changed, it replaces the ask instead of repeating it.
•Schedules stay declared enabled: false until a person reviews the new agent.

Headless websites

/create-agent:website runs the same way inside a deployed agent, such as a website-builder agent. The brief carries the answers (pages default to Home, About and Contact), decorative choices take defaults, and a real gap - usually the brand name or what the business does - goes out as one ask_operator question. Prices, addresses and phone numbers are never invented: the section is left out or marked TBD and listed in the summary. Each site is its own git repo under ~/sites/<project>, and re-running on an existing site edits it rather than re-scaffolding. Delivery uses keys the agent already holds, never an interactive login: a private GitHub repo through GITHUB_TOKEN or the platform git credential (the token is read at push time, never written into the remote URL), and a Vercel deploy through VERCEL_TOKEN. A missing key does not stop the build; the summary names the key a person should add as a credential in the agent's settings.

Usage Examples

Create an Agent for Any Domain

/create-agent:custom

The interview asks about:

•Role and purpose
•Skills and workflows
•Recommended schedules
•Credentials the agent needs
•Trinity wiring (template.yaml, .mcp.json.template, reporting)

Describe a sales researcher, a content writer, a knowledge base, or anything else — the same interview builds it.

Review and Improve an Existing Agent

/create-agent:review    # Read-only audit — prioritized findings, no changes
/create-agent:adjust    # Apply the improvements

review audits the agent against best practices - CLAUDE.md, skills, composition integrity, Trinity readiness - and produces a prioritized findings report without changing anything. Prose inter-agent delegation and prose schedule messages are reported as findings, as are a template.yaml with no plugins: block and a report guard that does not handle the refusal for a key without an agent identity. Deprecated paths are warnings: appending to the operator-queue file instead of raising an ask with ask_operator, and addressing a report or message by email instead of by role (to). adjust is the write-side companion: it proposes exact before/after changes and applies the ones you approve.

See Also

•agent-dev Plugin - Extend agents with skills and memory
•trinity Plugin - Deploy to Trinity platform
•Abilities Overview - Full toolkit overview