Subagent Prompt Foundry

A skill that generates deterministic, production-grade system prompts for custom subagents.

What It Does

Building a good subagent prompt is not the same as writing a regular instruction. A subagent runs outside the parent agent's context window — it cannot see the conversation history, the parent's assumptions, or anything not explicitly passed to it. This skill enforces a structured workflow: it researches the domain, asks targeted clarifying questions only when needed, drafts a prompt with explicit role, inputs, methodology, and output format, then validates it with bundled scripts before returning the final result. The output is a reusable system prompt artifact, not a one-off answer.

When To Use This

• "I need a system prompt for a code review subagent"
• "Make this prompt reusable across projects"
• "Design a subagent my parent agent can delegate security audits to"
• "Turn this one-off prompt into something production-grade"
• "Create a research subagent with strict source requirements"
• "Build a planning subagent that returns structured JSON to my parent agent"
• "I want a validator subagent that checks outputs before they go to the user"
• "Generate a system prompt for an agent that audits API specs"

When NOT To Use This

• You just want the task done directly ("review this code" → use a code review tool)
• You want generic prompting advice
• You want a meta-agent that writes prompts for other agents
• You just need the final answer, not a reusable artifact

Install

Any agent via skills.sh

npx skills add srinitude/subagent-prompt-foundry

Manual install (any agent)

Copy skills/subagent-prompt-foundry/ into your agent's skill directory.

Agent-specific paths

Agent	Path
Claude Code	.claude/skills/subagent-prompt-foundry/
OpenAI Codex	.agents/skills/subagent-prompt-foundry/
Cursor	.cursor/skills/subagent-prompt-foundry/
GitHub Copilot	.copilot/skills/subagent-prompt-foundry/
Gemini CLI	.gemini/skills/subagent-prompt-foundry/
Factory.ai	.factory/skills/subagent-prompt-foundry/
OpenCode	.opencode/skills/subagent-prompt-foundry/
Windsurf	.windsurf/skills/subagent-prompt-foundry/
Amplify	.amplify/skills/subagent-prompt-foundry/
Aider	.aider/skills/subagent-prompt-foundry/
Cline	.cline/skills/subagent-prompt-foundry/
Roo Code	.roo/skills/subagent-prompt-foundry/
Bolt	.bolt/skills/subagent-prompt-foundry/
Void	.void/skills/subagent-prompt-foundry/
Trae	.trae/skills/subagent-prompt-foundry/

Each agent directory contains a relative symlink to skills/subagent-prompt-foundry/, so there is only one copy of the skill content in the repo.

How It Works

1. Routes the request — checks whether this is actually a subagent prompt task (not direct execution, not generic advice)
2. Researches the domain — mandatory web research using first-party and official sources before drafting anything
3. Asks targeted clarifying questions — only if research did not resolve material ambiguity (role, inputs, deliverable, format, constraints)
4. Drafts the prompt — explicit role opener, bound inputs, step-by-step methodology, output schema, parent↔subagent protocol
5. Validates with bundled scripts — shape check confirms required sections exist; determinism scorer quantifies structural quality
6. Critiques against failure modes — self-critique pass catches generic-assistant drift, domain flattening, fake certainty
7. Returns only the finished system prompt — no explanation, no commentary unless explicitly requested

What Makes This Different

For agents evaluating whether to use this skill:

• Deterministic structure — same input class produces the same prompt skeleton, not random variations
• Mandatory research — never drafts from assumptions alone; always checks first-party sources before writing
• Isolated execution model — prompts define a parent↔subagent contract (what the parent sends, what the subagent returns), not just a list of instructions
• Script-validated — shape validation, determinism scoring, and version diffing are code, not informal checks
• Failure-mode aware — self-critique pass explicitly tests for the most common structural prompt failures

What's Inside

skills/subagent-prompt-foundry/
├── SKILL.md                          # Core instructions (401 lines)
├── agents/
│   └── openai.yaml                   # Codex-specific metadata
├── references/
│   ├── ROUTING_MATRIX.md             # When to activate vs. defer
│   ├── SUBAGENT_PROTOCOL.md          # Parent↔subagent message contract
│   ├── OUTPUT_STRUCTURES.md          # Default output templates
│   ├── FAILURE_MODES.md              # Critic-pass checklist
│   └── RESEARCH_AND_CLARIFICATION.md # Research and question policy
├── assets/
│   ├── subagent-template.md          # Starter skeleton for generated prompts
│   ├── routing-checklist.md          # Quick activation gate
│   └── parent-subagent-message-template.json # Handoff JSON schema
├── scripts/
│   ├── validate_prompt_shape.py      # Checks required sections exist
│   ├── score_determinism.py          # Scores structural determinism 0-1
│   └── diff_prompt_versions.py       # Unified diff between prompt versions
└── evals/
    └── evals.json                    # 12 routing test cases

Requirements

• Python 3.6+ (for bundled validation scripts)
• Network access (for mandatory web research step)

License

MIT

Author

Kiren Srinivasan