---

name: metaprompt-factory description: This skill should be used when generating structured metaprompts for repeatable tasks. Triggers include "create a metaprompt for X", "generate a review prompt", "make a prompt template", or when building prompts for copilot, Codex, or other AI tools.

metaprompt-factory

generate structured XML metaprompts for repeatable AI tasks. adapts to domain, enforces quality gates, cites sources.

philosophy

"a metaprompt is a reusable product spec for an AI worker"

when to use

decision tree: prompt structure

CopyWhat kind of task?
├── Classification/Categorization
│   ├── Use multishot with 3-5 diverse examples
│   ├── Wrap examples in <example></example> tags
│   └── Include edge cases in examples
├── Analysis (code review, document review)
│   ├── Use XML to separate <document> from <instructions>
│   ├── Request structured output with specific sections
│   └── Add <output_format> tag with example structure
├── Generation (content, code)
│   ├── Define role/persona in system prompt
│   ├── Provide <formatting_example> for style
│   └── Use prefill to guide output structure
├── Multi-step reasoning
│   ├── Use chain-of-thought: <thinking> then <answer>
│   ├── Break into subtasks, chain prompts
│   └── Consider self-correction loop
└── Validation/Review
    ├── Generate → Review → Refine pattern
    ├── Use graded rubric (A-F or 1-10)
    └── Request specific improvement suggestions

decision tree: complexity level

CopyHow complex is the task?
├── Simple (single transformation)
│   ├── Zero-shot or 1 example sufficient
│   ├── Direct instructions, minimal structure
│   └── ~50-100 tokens prompt
├── Moderate (multiple considerations)
│   ├── 3-5 multishot examples
│   ├── XML structure for clarity
│   ├── Explicit output format
│   └── ~200-500 tokens prompt
├── Complex (multi-step, high-stakes)
│   ├── Chain prompts (break into subtasks)
│   ├── Self-correction loop (generate → review → refine)
│   ├── Comprehensive examples with edge cases
│   └── ~500-1000+ tokens prompt
└── Autonomous (agent workflows)
    ├── Full XML structure with all sections
    ├── Failure modes documented
    ├── Verification steps included
    └── Tool integration specified

concrete patterns (from Anthropic docs)

XML tag patterns

multishot guidelines

chain prompt patterns

workflow

1. intake

gather minimal viable inputs:

Copytarget_task: what the prompt should accomplish
domain: coding | review | ops | research | classification
outputs: expected deliverables (structured format)
tooling: copilot | codex | claude-api

2. classify complexity

use decision tree above to determine:

• simple → direct prompt
• moderate → XML structure + multishot
• complex → chained prompts
• autonomous → full metaprompt

3. select pattern

4. structure metaprompt

Copy<?xml version="1.0" encoding="UTF-8"?>
<metaprompt name="{name}" version="1.0">
  <role>
    <!-- who is the AI in this context? -->
  </role>

  <context>
    <!-- background, constraints, assumptions -->
  </context>

  <instructions>
    <!-- numbered steps, clear directives -->
  </instructions>

  <examples>
    <example>
      <input>...</input>
      <output>...</output>
    </example>
    <!-- 3-5 diverse examples -->
  </examples>

  <output_format>
    <!-- exact structure expected -->
  </output_format>

  <failure_modes>
    <!-- what can go wrong, how to handle -->
  </failure_modes>
</metaprompt>

5. parameterize

extract variables for reuse:

Copy<parameters>
  <param name="document" type="string" required="true">
    The document to analyze
  </param>
  <param name="format" type="enum" values="json,yaml,text" default="json">
    Output format
  </param>
</parameters>

6. validate

check against rubric:

tool integration

trails integration

persist prompt creation for pattern analysis:

Copy# record prompt creation
trails trail record --agent claude --action completed \
  --task "metaprompt-factory: $PROMPT_NAME - $DOMAIN" \
  --confidence $CONFIDENCE --json -q

trails enables:

• tracking prompt patterns across domains
• measuring prompt quality over time
• correlating prompts with execution success

copilot (consult-light)

Copycat <<'EOF' | copilot -p --model gemini-3-pro
<context>
{context_packet}
</context>

<instructions>
1. Analyze the provided context
2. Output structured JSON
</instructions>

<output_format>
{field: type, ...}
</output_format>
EOF

codex (consult-deep)

Copycat <<'EOF' | codex exec --model "gpt-5.2-codex xhigh"
<role>
Senior code reviewer with expertise in {domain}
</role>

<document>
{code_or_document}
</document>

<instructions>
1. Review for {criteria}
2. Provide actionable feedback
3. Rate severity (high/medium/low)
</instructions>

<output_format>
{
  "issues": [...],
  "summary": "...",
  "score": 1-10
}
</output_format>
EOF

anti-patterns

quality rubric

scoring:

• 9-10: production-ready, comprehensive examples, handles edge cases
• 7-8: solid, minor gaps in examples or failure modes
• 5-6: functional but needs more examples or structure
• <5: needs significant rework

references

• references/xml-patterns.md - complete XML tag reference with examples
• references/domain-templates.md - templates for coding, review, ops, classification
• references/chain-patterns.md - multi-step prompt chaining patterns
• references/domain-modules.md - domain-specific prompt modules
• references/rubric.md - quality scoring rubric
• references/factory-template.xml - base metaprompt template

examples

classification prompt

Copy<role>
Customer feedback analyst for a B2B SaaS product.
</role>

<instructions>
Analyze feedback and categorize issues. Use categories:
UI/UX, Performance, Feature Request, Integration, Pricing, Other.
Rate sentiment (Positive/Neutral/Negative) and priority (High/Medium/Low).
</instructions>

<examples>
<example>
<input>The new dashboard is a mess! It takes forever to load, and I can't find the export button. Fix this ASAP!</input>
<output>
Category: UI/UX, Performance
Sentiment: Negative
Priority: High
</output>
</example>
<example>
<input>Love the Salesforce integration! But it'd be great if you could add Hubspot too.</input>
<output>
Category: Integration, Feature Request
Sentiment: Positive
Priority: Medium
</output>
</example>
</examples>

<data>
{{FEEDBACK}}
</data>

code review prompt

Copy<role>
Senior engineer reviewing pull request for security and performance.
</role>

<document>
{{PR_DIFF}}
</document>

<instructions>
1. Identify security vulnerabilities (OWASP top 10)
2. Flag performance concerns (N+1 queries, memory leaks)
3. Check for test coverage gaps
4. Rate each issue: severity (high/medium/low), effort (small/medium/large)
</instructions>

<output_format>
{
  "security_issues": [{"description": "", "severity": "", "line": 0}],
  "performance_issues": [{"description": "", "severity": "", "line": 0}],
  "test_gaps": [""],
  "overall_rating": "approve | request_changes | comment",
  "summary": ""
}
</output_format>