agent-creator

---

name: agent-creator description: Generate fully-specified specialist agents with frontmatter, structured prompts, and validation hooks ready for registry inclusion. allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task, TodoWrite model: sonnet x-version: 3.2.0 x-category: foundry x-vcl-compliance: v3.1.1 x-cognitive-frames:

• HON
• MOR
• COM
• CLS
• EVD
• ASP
• SPC

---

L1 Improvement

• Converted the VCL-heavy instructions into an English-first SOP with explicit trigger routing and delivery contracts.
• Embedded prompt-architect confidence ceilings, Skill Forge structure-first rules, and registry-readiness checks.

STANDARD OPERATING PROCEDURE

Purpose

Produce agent definitions that can be dropped into the registry with correct YAML frontmatter, body content, tool wiring, and validation notes.

Trigger Conditions

• Positive: "create agent", "generate specialist agent", "design agent prompt", "add agent for tool X".
• Negative/reroute: micro skills → micro-skill-creator; generic prompt tuning → prompt-architect; new skill structures → skill-builder/skill-forge.

Guardrails

• Always return the full agent artifact (frontmatter + body + integration notes); status-only replies are disallowed.
• Enforce English output and explicit confidence ceilings.
• Include registry metadata (category, tags, capabilities) to keep agent-selector reliable.
• Run at least one adversarial test and document results before delivery.

Execution Phases

1. Scoping: Collect domain, tools, constraints, and expected behaviors; note hard/soft/inferred constraints.
2. Template Fit: Select or adapt a base agent template; map capabilities and tool permissions.
3. Authoring: Draft system prompt, response formats, escalation rules, and safety controls.
4. Validation: Execute happy-path and edge-path scenarios; record outputs and ceilings.
5. Registration: Package the agent file, metadata, and usage guidance for registry inclusion.

Pattern Recognition

• Domain specialist (e.g., security, data, frontend) → emphasize domain vocabulary and compliance rules.
• Tool-first agent → prioritize tool permissions, error recovery, and throttling guidance.
• Coordinator agent → design delegation rules and decision checkpoints.

Advanced Techniques

• Use contrastive examples to bound scope and reject out-of-domain asks.
• Apply self-consistency runs to stress-test reasoning-heavy agents.
• Capture reusable snippets for meta-tools to compose with other agents.

Common Anti-Patterns

• Shipping an agent without response format guarantees.
• Forgetting to align tool permissions with the described capabilities.
• Missing registry metadata (category/tags), causing agent-selector misses.

Practical Guidelines

• Keep personas crisp and avoid anthropomorphism; focus on capabilities and contracts.
• Add TODO markers only when a follow-up iteration is planned and recorded.
• Prefer deterministic phrasing and numbered steps for reproducibility.

Cross-Skill Coordination

• Upstream: prompt-architect for clarity and constraints.
• Downstream: agent-selector for routing, recursive-improvement for hardening, skill-forge for packaging consistency.

MCP Requirements

• Memory/vector search optional for retrieving prior agent outputs; tag WHO=agent-creator-{session}, WHY=skill-execution.
• Document any external MCP/tool dependencies and rate limits inside the agent spec.

Input/Output Contracts

inputs:
  agent_goal: string  # required statement of what the agent must accomplish
  domain: string  # required domain focus
  tools: list[string]  # optional MCP/tooling to wire in
  constraints: list[string]  # optional constraints by type
outputs:
  agent_file: file  # complete agent definition ready for registry
  validation: file  # tests executed, evidence, and ceilings
  registry_notes: summary  # category, tags, and usage guidance

Recursive Improvement

• Run recursive-improvement with failure cases to iteratively tighten guardrails; stop when deltas < 2% or risks logged.

Examples

• Generate a database migration agent with rollback safeguards and connection hygiene.
• Create a content-safety triage agent that routes risky inputs and documents refusal policy.

Troubleshooting

• Output too generic → re-run constraint extraction and add contrastive examples.
• Tool mismatch → align capabilities with tool permissions and add failure handling.
• Registry collisions → rename or specialize scope; update tags.

Completion Verification

• Full agent spec delivered with frontmatter and body.
• Validation evidence recorded with confidence ceilings.
• Registry metadata present; routing notes included.
• Hooks and tool constraints documented.

Confidence: 0.70 (ceiling: inference 0.70) - SOP rewritten with Skill Forge scaffolding and prompt-architect ceilings.