Prompt Manager — Quick Reference

Expert-only knowledge for transforming vague prompts into optimized, actionable ones. This guide focuses on non-obvious patterns, hard-won insights, and systematic optimization techniques.

How to Use This Guide

1. Start with SKILL.md — Run the 4-phase workflow for any prompt optimization
2. Reference this guide — Quick lookup for patterns, frameworks, and techniques
3. Load details on-demand — Deep-dive into specific references/ files as needed

Quick Reference

Anti-Patterns (NEVER Do)

Fabrication Techniques — MoE/ToT/GoT make Claude invent fake personas instead of deepening reasoning → references/credit-killing-patterns.md#fabrication

CoT on Reasoning Models — Claude 4.5+ has native extended thinking, adding "think step-by-step" degrades quality → references/credit-killing-patterns.md#cot

Framework Name Pollution — Never output "Using CO-STAR..." or label sections with methodology → references/credit-killing-patterns.md#naming

Context-Free Optimization — Prompts for Claude Code differ from ChatGPT/API calls, context determines strategy → references/credit-killing-patterns.md#context

Vague Success Criteria — "Better", "comprehensive", "clean" lack measurability, pin to objective outcomes → references/credit-killing-patterns.md#criteria

Missing Creative Constraints — Without boundaries, creative tasks produce inconsistent results → references/credit-killing-patterns.md#constraints

Lost-in-the-Middle — Models weaken attention on middle sections of long prompts, place critical info at start/end → references/credit-killing-patterns.md#context-positioning

Ambiguous Pronouns — "It/this/that" become unclear in multi-step workflows, use specific nouns → references/credit-killing-patterns.md#pronouns

Frameworks

When to Use Which:

• CO-STAR → Structured output, specific format needs (format-driven)
• RISEN → Multi-step procedures, workflows (process-driven)
• RODES → Needs examples for clarity, style matching (example-driven)

Application Rule: Route user intent through framework structure silently — never expose methodology in output

→ Full details: references/frameworks.md

Optimization Techniques

5 Safe Techniques that improve prompt quality without risk:

1. Specificity Injection — Replace vague terms with concrete criteria
2. Constraint Addition — Define boundaries for creative freedom
3. Context Positioning — Critical info at start/end, not middle
4. Pronoun Elimination — Replace "it/this/that" with specific nouns
5. Success Criteria Definition — Pin to measurable outcomes

→ Detailed examples: references/safe-techniques.md

Complexity Assessment

Simple: Single objective, <3 steps, no ambiguity → Execute directly Moderate: Some ambiguity, 3-5 steps, few dependencies → Proceed with monitoring Complex: >3 interdependent decisions OR >5 sequential phases → Recommend plan mode

→ Full criteria: references/complexity-detection.md

Plan Mode Triggers

Always recommend plan mode when:

• 3 interdependent decisions that affect each other

• 5 sequential phases requiring coordination

• Significant ambiguity with cascading implications
• User asks "how should I approach this?"

→ Detailed triggers: references/plan-mode-triggers.md

Ambiguity Patterns

Common Vague Terms:

• "Comprehensive" → All edge cases [+time] vs common scenarios [balanced] vs overview [+speed]
• "Fast" → Response time, development time, or execution time?
• "Simple" → Minimal code, easy to understand, or few dependencies?
• "Clean" → Follows standards, minimal complexity, or well-documented?

Resolution Pattern: Present 2-3 interpretation options with implications, let user decide

→ Full catalog: references/ambiguity-examples.md

Template Selection

12 Task-Type Templates:

• Analytical — Data analysis, metrics, research
• Creative — Writing, design, ideation
• Debugging — Error investigation, root cause analysis
• Documentation — README, API docs, guides
• Exploration — Codebase discovery, pattern finding
• Implementation — Feature building, coding
• Planning — Architecture, approach design
• Refactoring — Code improvement, restructuring
• Review — Code review, quality assessment
• Security — Vulnerability analysis, threat modeling
• Testing — Test writing, QA, validation
• Troubleshooting — System issues, incident response

→ Selection logic: references/template-selection.md → Full templates: assets/prompt-templates/

Optimization Examples

Before/After Transformations:

• Vague request → Specific, constrained prompt
• Ambiguous multi-step → Clear sequential workflow
• Missing context → Fully specified execution environment
• Generic goal → Measurable success criteria

→ Full examples: references/optimization-examples.md

Workflow Summary

Phase 1: Intake & Assessment
  ↓ Extract intent, calibrate skill level, detect complexity
Phase 2: Pattern Detection
  ↓ Identify credit-killing patterns, ambiguities, trade-offs
Phase 3: Framework Selection & Optimization
  ↓ Apply CO-STAR/RISEN/RODES silently, use safe techniques
Phase 4: Validation & Handoff
  ↓ Quality checks, flag remaining ambiguities, recommend execution mode

Important Context

Model-Specific: Claude 4.5+ uses native extended thinking, GPT-4 uses internal CoT — optimization strategies differ by model family

Token Economy: System prompts multiply by API call count — progressive disclosure reduces base cost

Security: Validate user inputs, use delimiters, never allow user content to override system instructions

Memory Persistence: Save optimization patterns to memory blocks to prevent contradictions in extended conversations