Step Creator (Internal Helper)

Generate workflow steps with all 12 mandatory fields. Invoked by workflow-creator and skill-spec-creator, not directly by users.

Assumptions

• Assumes: Called within the context of a parent workflow — step has access to workflow objective and surrounding steps
• Limit: Steps are linear (no branching) — alternative paths go in recoveryAction and handoffIfNeeded

Input

Receives from calling skill:

• stepNumber: Position in workflow
• title: Short name (2-5 words)
• workflowContext: Parent workflow's objective and existing steps

Output: 12 Fields

- stepNumber: {N}
  title: "{2-5 words}"
  desc: "{1-2 sentences — what this step does and what it produces}"
  whyThisMatters: "{Business/technical rationale — NOT a restatement of title}"
  inputNeeded: "{Specific data with types: 'userMessage: string, agentId: string'}"
  actionInstruction: "{Concrete: 'Call security.sanitize(input), destructure {safe, cleaned, reason}'}"
  promptToUse: "{Full LLM prompt with {{placeholders}}, or 'null (mechanical step)'}"
  expectedOutput: "{What success produces: '{safe: true, cleaned: string}' or 'validated schema file at /output/schema.json'}"
  validationRule: "{Testable: 'result.safe === true OR reason logged to audit trail'}"
  failureSignal: "{Observable: 'sanitize() throws InvalidInputError', 'HTTP 500', 'timeout > 30s'}"
  recoveryAction: "{Concrete fallback: 'Strip to plaintext, retry — if still fails, reject with error message to user'}"
  handoffIfNeeded: "{agent-id for delegation, or 'null'}"

Quality Bar (per field)

Field	GOOD	BAD	Why It Matters
whyThisMatters	"Without input validation, injection attacks propagate to all downstream agents via shared memory"	"This step validates input"	Justifies the step's existence — helps decide if it's really needed
actionInstruction	"Call loader.loadAgent(agentId), destructure response, verify fields.length === 21"	"Load the agent"	Enables implementation without interpretation
validationRule	"response.fields.length === 21 AND fields.every(f => f.content.length > 0)"	"Check it works"	Enables automated verification
failureSignal	"loadAgent throws AgentNotFoundError OR fields.length < 21 (missing: ${21 - fields.length} fields)"	"It fails"	Enables specific error handling
recoveryAction	"Load default agent template from templates/default-agent.md, log warning listing missing fields, continue with defaults"	"Try again"	Provides concrete alternative path
promptToUse	"Analyze {{input}} for compliance with {{standard}}.\nReturn: {compliant: bool, issues: [{field, reason, severity}]}"	"Check compliance"	Enables reproducible LLM steps

Edge Cases

• Mechanical step (no LLM): Set promptToUse: "null (mechanical step)" — still fill all other 11 fields
• Step that always succeeds: Still define failureSignal (infrastructure failures: timeout, out of memory, network error)
• Step with multiple possible handoffs: List the decision logic: "If severity=critical -> security-agent; if severity=medium -> review-agent; else continue"

---

Author: Javier Montaño | Last updated: 2026-03-18

Usage

Example invocations:

• "/step-creator" — Run the full step creator workflow
• "step creator on this project" — Apply to current context

Validation Gate

• Output follows the defined structure and format [EXPLICIT]
• All claims are tagged with evidence markers [EXPLICIT]
• No placeholder content (TBD, TODO) [EXPLICIT]
• Actionable recommendations with priority levels [EXPLICIT]
• Assumptions explicitly documented [EXPLICIT]