Prompting

Generates, optimizes, and composes prompts programmatically. It's the standard library for prompt engineering — other skills call it when they need to build or improve a prompt. The output is always a prompt to be used elsewhere, never the final content itself.

Invoke when: meta-prompting, template generation, prompt optimization, programmatic prompt composition, creating dynamic agents, generating structured prompts from data.

Prompt engineering tends to get copy-pasted and rewritten by hand across every skill that needs it, so the same patterns drift apart and best practices live in one person's head. When you want to compose a prompt from data — spin up a custom agent, build an eval judge, generate a phased workflow — there's no clean way to separate the structure from the content. This skill makes structure code and content data: one Handlebars template plus different data renders specialized agents, workflows, and eval frameworks, and the engineering standards live in one place every skill can reference.

Three pillars carry the work:

• Standards - Anthropic best practices, Claude 4.x patterns, empirical research (markdown-first design, context engineering, the Fabric pattern system, 1,500+ academic papers on prompt optimization). Full guide in Standards.md.
• Templates - Handlebars-based system for programmatic prompt generation: Briefing, Structure, Gate, DynamicAgent, plus eval templates (Judge, Rubric, TestCase, Comparison, Report).
• Tools - Template rendering (RenderTemplate.ts), validation, and data-content separation.

Example 1: Using Briefing Template (Agent Skill)

// skills/Agents/Tools/ComposeAgent.ts
import { renderTemplate } from '${CLAUDE_SKILL_DIR}/Tools/RenderTemplate.ts';

const prompt = renderTemplate('Primitives/Briefing.hbs', {
  briefing: { type: 'research' },
  agent: { id: 'EN-1', name: 'Skeptical Thinker', personality: {...} },
  task: { description: 'Analyze security architecture', questions: [...] },
  output_format: { type: 'markdown' }
});

Example 2: Using Structure Template (Workflow)

# Data: phased-analysis.yaml
phases:
  - name: Discovery
    purpose: Identify attack surface
    steps:
      - action: Map entry points
        instructions: List all external interfaces...
  - name: Analysis
    purpose: Assess vulnerabilities
    steps:
      - action: Test boundaries
        instructions: Probe each entry point...

bun run RenderTemplate.ts \
  --template Primitives/Structure.hbs \
  --data phased-analysis.yaml

Example 3: Custom Agent with Voice Mapping

// Generate specialized agent with appropriate voice
const agent = composeAgent(['security', 'skeptical', 'thorough'], task, traits);
// Returns: { name, traits, voice: 'default', voiceId: 'VOICE_ID...' }

Agents Skill

• Uses Templates/Primitives/Briefing.hbs for agent context handoff
• Uses RenderTemplate.ts to compose dynamic agents
• Maintains agent-specific template: Agents/Templates/DynamicAgent.hbs

Evals Skill

• Uses eval-specific templates: Judge, Rubric, TestCase, Comparison, Report
• Leverages RenderTemplate.ts for eval prompt generation
• Eval templates may be stored in Evals/Templates/ but use Prompting's engine

Development Skill

• References Standards.md for prompt best practices
• Uses Structure.hbs for workflow patterns
• Applies Gate.hbs for validation checklists

The templating system eliminated ~35,000 tokens (65% reduction) across PAI:

1. Separation of Concerns

• Templates: Structure and formatting only
• Data: Content and parameters (YAML/JSON)
• Logic: Rendering and validation (TypeScript)

2. Keep Templates Simple

• Avoid complex logic in templates
• Use Handlebars helpers for transformations
• Business logic belongs in TypeScript, not templates

3. DRY Principle

• Extract repeated patterns into partials
• Use presets for common configurations
• Single source of truth for definitions

4. Version Control

• Templates and data in separate files
• Track changes independently
• Enable A/B testing of structures

Primary Documentation:

• Standards.md - Complete prompt engineering guide
• Templates/README.md - Template system overview (if preserved)
• Tools/RenderTemplate.ts - Implementation details

Research Foundation:

• Anthropic: "Claude 4.x Best Practices" (November 2025)
• Anthropic: "Effective Context Engineering for AI Agents"
• Anthropic: "Prompt Templates and Variables"
• The Fabric System (January 2024)
• "The Prompt Report" - arXiv:2406.06608
• "The Prompt Canvas" - arXiv:2412.05127

Related Skills:

• Agents - Dynamic agent composition
• Evals - LLM-as-Judge prompting
• Development - Spec-driven development patterns

Philosophy: Prompts that write prompts. Structure is code, content is data. Meta-prompting enables dynamic composition where the same template with different data generates specialized agents, workflows, and evaluation frameworks. This is core PAI DNA - programmatic prompt generation at scale.

• Meta-prompting generates PROMPTS, not content. The output is a prompt that gets used elsewhere — not the final deliverable.
• Templates should be model-agnostic. Don't write prompts that depend on specific model quirks.
• Test generated prompts before declaring them ready. A prompt that looks good may perform poorly.