Obtain a workflow script

Available now The prompt and skill use the published workflow-authoring guide.

Use the guide that matches the released workflow engine. Do not tell an agent to guess the workflow domain-specific language (DSL).

Use the MCP prompt

Open the author-workflow prompt in an MCP host that supports prompts. Give the prompt a clear request.

Use this request for the first workflow:

Author a read-only AgentPrism workflow that inspects the current project for one request. Return a structured summary, supporting file paths, and one next step. Use Codex with gpt-5.6-sol, read-only mode, and reasoning_effort: "max". Save the workflow script as first-inspection.workflow.js. Before the first live run, validate the workflow script.

The host agent must return plain JavaScript. The first statement must be the meta export.

Use the published skill

If the host does not support MCP prompts, use this method.

Install the published authoring skill:

Published authoring skill · installer command

npx skills add agentprism/agentprism-workflows

Tell the host agent to use agentprism-workflow-authoring. Give it the same request from the prior section.

Use the verified first workflow

You can use this workflow without authoring a new one. Save it as first-inspection.workflow.js in your project.

first-inspection.workflow.js · workflow script

export const meta = {
  name: 'first-inspection',
  description: 'Inspect one project and return a structured result',
  phases: [{ title: 'Inspect' }],
};

const input =
  args && typeof args === 'object' && !Array.isArray(args) ? args : {};
if (typeof input.request !== 'string' || !input.request.trim()) {
  throw new Error('args.request must be a non-empty string');
}
const request = input.request.trim();

const RESULT_SCHEMA = {
  type: 'object',
  additionalProperties: false,
  required: ['summary', 'files', 'nextStep'],
  properties: {
    summary: {
      type: 'string',
      description: 'A short answer with evidence from the project',
    },
    files: {
      type: 'array',
      description: 'Project files that support the answer',
      items: { type: 'string' },
    },
    nextStep: {
      type: 'string',
      description: 'The next recommended action',
    },
  },
};

phase('Inspect');
const result = await agent(
  `Inspect the current project for this request:\n${request}\n` +
    'Do not change files. Return evidence from the project.',
  {
    label: 'inspect-project',
    model: 'codex/gpt-5.6-sol',
    mode: 'read-only',
    configOptions: { reasoning_effort: 'max' },
    schema: RESULT_SCHEMA,
    retries: 1,
  },
);

if (!result) throw new Error('The inspection call returned no result');
return result;

The script has three parts: the meta export, one read-only agent() call routed to codex/gpt-5.6-sol, and a structured return value.

The schema requires all properties and sets additionalProperties: false.

Review the workflow script

Before you validate the workflow script, make sure that it has these properties:

• The first statement is export const meta = { ... }.
• The workflow script has no imports.
• The workflow script captures and validates args.request before the first call.
• The call has a stable label.
• The call has an explicit model.
• The schema uses a root object.
• The workflow script does not use Date.now(), new Date(), or Math.random().

Next step

Validate the workflow script.