managing-prompts

Managing Prompts

Comprehensive prompt engineering for Claude 4.5: creation, analysis, optimization, guardrails, context management.

When to Use

• User requests creating/analyzing/optimizing prompts
• User asks if current logic should be extracted to prompt
• User mentions: hallucinations, consistency issues, context optimization, prompt caching, chain-of-thought, XML tags
• Context usage high and extraction would help
• Outdated prompt techniques need Claude 4.5 updates

Core Workflows

Workflow 1: Analyzing Existing Prompts

Quick process:

1. Read prompt file/content
2. Load quick-reference.md for validation checklist
3. Apply evaluation: structure, techniques, guardrails, context efficiency, Claude 4.5 optimization
4. Categorize issues: critical | major | minor
5. Generate recommendations with priority
6. Output using output-formats.md#analysis-report template

Detailed evaluation: Load analysis-patterns.md

Common issues reference: quick-reference.md#common-issues

---

Workflow 2: Creating New Prompts

Quick process:

1. Gather requirements: objective, success criteria, inputs/outputs, complexity, risk factors
2. Select architecture: Load architecture-patterns.md, choose: simple_task | complex_reasoning | agent_workflow | multi_window
3. Apply techniques: Load technique-reference.md for indexed lookup by need
4. Add guardrails: Based on risk level, load guardrails-implementation.md for patterns
5. Optimize context: Load context-optimization.md if >10K tokens
6. Validate: Use quick-reference.md#quick-validation-checklist

Technique order:

1. Clear instructions (explicit, contextual motivation, success criteria, output format)
2. Examples (3-5 for complex tasks)
3. Chain-of-thought (for reasoning tasks)
4. XML tags (for structure)
5. Role assignment, prefilling, template variables

Architecture selection:

• Simple task: Clear objective, straightforward
• Complex reasoning: Multi-step analysis, judgment
• Agent workflow: Tool access, iterative, autonomous
• Multi-window: >100K tokens, extended work

Guardrails by risk:

• Low: Basic uncertainty permission
• Medium: Examples + format spec
• High: Quote-grounding + citations
• Critical: Full stack (quotes, citations, knowledge restriction, CoT)

Context optimization:

• Cache static content ≥1024 tokens
• Place 20K+ docs at beginning
• Remove Claude's existing knowledge
• Quote-grounding for large docs
• Split if >3 distinct subtasks

---

Workflow 3: Optimizing Existing Prompts

Quick process:

1. Identify goal: reduce_tokens | improve_quality | increase_consistency | reduce_hallucinations | enable_caching | split_prompt
2. Load optimization-strategies.md for before/after examples
3. Apply transformation
4. Measure improvement
5. Document using output-formats.md#optimization-report

Quick optimization strategies:

• Context reduction: Remove redundancy, cache static, quote-grounding, split if >3 subtasks
• Quality: Add examples, CoT, explicit instructions, contextual motivation
• Consistency: Format spec, prefilling, 3-5 examples
• Hallucinations: Uncertainty permission, quote-grounding, citations
• Caching: Identify ≥1024 static tokens, restructure, add cache_control

Decision aids: quick-reference.md#fast-decision-trees

---

Workflow 4: Updating Outdated Prompts

Quick process:

1. Load migration-guide.md for Claude 3→4.5 patterns
2. Identify outdated: insufficient explicitness, missing motivation, suggestion framing, no summary requests
3. Apply updates preserving core logic
4. Test and document

Common migrations:

• Vague → Explicit instructions with specifics
• No motivation → Add contextual reasoning
• Suggestion framing → Action framing ("Improve this" not "Can you improve?")
• No summaries → Explicit requests (4.5 less verbose by default)
• No quality modifiers → Add "Give it your all" for creative tasks

---

Workflow 5: Deciding Logic Extraction

Quick process:

1. Load extraction-decision-guide.md#decision-tree
2. Evaluate criteria (6 for prompt, 6 for script)
3. Score: 3+ = extract, <3 = keep inline
4. Output using output-formats.md#extraction-decision template

Prompt extraction criteria (3+ to extract):

• Reused across 3+ contexts
• Needs prompt techniques (examples, CoT, XML)
• Benefits from guardrails
• Context optimization beneficial
• Standardization needed
• Complex (>200 tokens instructions)

Script extraction criteria (3+ to extract):

• Deterministic/algorithmic
• Computation-heavy
• Unit testable
• Reused across prompts
• No reasoning required
• Format conversion/validation

Keep inline: One-time, context-specific, simple, requires real-time interaction

Fast Reference System

For validation: quick-reference.md (quick checklist, common issues, decision trees, severity guide)

For technique lookup: technique-reference.md (indexed by use case, lookup table, application patterns, combinations)

For output templates: output-formats.md (analysis, optimization, extraction, creation, migration reports)

For detailed patterns:

• extended-thinking-implementation.md (internal reasoning, budget management, batch processing, multi-round strategies)
• prompt-chaining-architecture.md (sequential/parallel workflows, handoff patterns, orchestration, self-correction loops)
• consistency-techniques.md (prefilling, Structured Outputs, format enforcement, advanced patterns)
• claude-4-5-optimization.md (4.5-specific guidance: explicit instructions, context awareness, multi-context workflows, tool usage)
• architecture-patterns.md (5 pattern types: simple task, complex reasoning, agent, multi-window, multi-context workflows)
• guardrails-implementation.md (hallucination, consistency, security, character patterns, jailbreak prevention, prompt leak reduction)
• context-optimization.md (token management, caching with model-specific minimums, document org, splitting, cache invalidation patterns)
• optimization-strategies.md (11 before/after examples)
• migration-guide.md (Claude 3→4.5 patterns, 10 migration examples)
• extraction-decision-guide.md (decision trees, criteria, examples)
• analysis-patterns.md (common issues, quality assessment, evaluation examples)

Claude 4.5 Critical Behaviors

Strengths: Long-horizon reasoning, context awareness, parallel execution, subagent delegation

Requirements: More explicit than Claude 3, contextual motivation, high-quality examples, action framing, explicit summary requests, quality modifiers

Optimization: Parallel tool guidance, clear success criteria, investigation before answering, aesthetic direction (UI), algorithm approach (code)

Quick Troubleshooting

Load quick-reference.md for detailed troubleshooting with fixes.

Common symptoms → Quick fixes:

• Inconsistent outputs → Format spec + 3-5 examples
• Hallucinations → Quote-grounding + uncertainty permission
• High tokens → Cache static content, remove redundancy
• Too concise (4.5) → Explicit detail requests
• Suggests not implements (4.5) → Action framing
• Stops early (4.5) → Reference token budget available

Tool Usage Pattern

Standard workflow:

1. Identify which workflow (1-5)
2. Load quick-reference.md for checklist/decisions
3. Load technique-reference.md if need technique lookup
4. Load detailed supporting file only when needed
5. Use output-formats.md template for deliverable

Example - Analysis:

1. Read prompt
2. quick-reference.md → validation checklist
3. technique-reference.md → identify techniques used
4. analysis-patterns.md → detailed framework (if complex)
5. output-formats.md#analysis-report → generate output

Example - Creation:

1. Gather requirements
2. technique-reference.md → select techniques by need
3. architecture-patterns.md → get template for pattern type
4. quick-reference.md → validation before finalizing
5. output-formats.md#creation-summary → document (optional)

Example - Optimization:

1. Identify goal
2. quick-reference.md#fast-decision-trees → decide strategy
3. optimization-strategies.md → get before/after pattern
4. Apply and test
5. output-formats.md#optimization-report → document

Validation Before Deployment

Load quick-reference.md#quick-validation-checklist

Essential checks:

• Objective clear, success criteria defined
• Instructions explicit (action framing for 4.5)
• Output format specified
• Examples if complex (3-5, diverse, identical structure)
• CoT if reasoning required
• XML if multiple components
• Guardrails match risk level
• Context optimized (caching, placement, splitting)
• Claude 4.5 specificity (explicit, motivation, quality modifiers)

Golden test: "Would colleague understand with minimal context?"

Summary

This skill provides prompt engineering optimized for Claude 4.5. Use workflows systematically, load references on-demand, validate thoroughly. Supporting files contain detailed patterns - SKILL.md provides navigation and quick process guidance.