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- Outdated prompt techniques need Claude 4.5 updates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:
- Read prompt file/content
- Load quick-reference.md for validation checklist
- Apply evaluation: structure, techniques, guardrails, context efficiency, Claude 4.5 optimization
- Categorize issues: critical | major | minor
- Generate recommendations with priority
- 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:
- Gather requirements: objective, success criteria, inputs/outputs, complexity, risk factors
- Select architecture: Load architecture-patterns.md, choose: simple_task | complex_reasoning | agent_workflow | multi_window
- Apply techniques: Load technique-reference.md for indexed lookup by need
- Add guardrails: Based on risk level, load guardrails-implementation.md for patterns
- Optimize context: Load context-optimization.md if >10K tokens
- Validate: Use quick-reference.md#quick-validation-checklist
Technique order:
- Clear instructions (explicit, contextual motivation, success criteria, output format)
- Examples (3-5 for complex tasks)
- Chain-of-thought (for reasoning tasks)
- XML tags (for structure)
- 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:
- Identify goal: reduce_tokens | improve_quality | increase_consistency | reduce_hallucinations | enable_caching | split_prompt
- Load optimization-strategies.md for before/after examples
- Apply transformation
- Measure improvement
- 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:
- Load migration-guide.md for Claude 3→4.5 patterns
- Identify outdated: insufficient explicitness, missing motivation, suggestion framing, no summary requests
- Apply updates preserving core logic
- 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:
- Load extraction-decision-guide.md#decision-tree
- Evaluate criteria (6 for prompt, 6 for script)
- Score: 3+ = extract, <3 = keep inline
- 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:
- Identify which workflow (1-5)
- Load quick-reference.md for checklist/decisions
- Load technique-reference.md if need technique lookup
- Load detailed supporting file only when needed
- Use output-formats.md template for deliverable
Example - Analysis:
- Read prompt
- quick-reference.md → validation checklist
- technique-reference.md → identify techniques used
- analysis-patterns.md → detailed framework (if complex)
- output-formats.md#analysis-report → generate output
Example - Creation:
- Gather requirements
- technique-reference.md → select techniques by need
- architecture-patterns.md → get template for pattern type
- quick-reference.md → validation before finalizing
- output-formats.md#creation-summary → document (optional)
Example - Optimization:
- Identify goal
- quick-reference.md#fast-decision-trees → decide strategy
- optimization-strategies.md → get before/after pattern
- Apply and test
- output-formats.md#optimization-report → document
Validation Before Deployment
Load quick-reference.md#quick-validation-checklist
Essential checks:
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.
1---2name: managing-prompts3description: Creates, analyzes, updates, and optimizes prompts using Claude 4.5 best practices, guardrails, context management, and prompt engineering patterns. Use when user asks how to write effective prompts, explaining prompt engineering techniques, understanding Claude 4.5 best practices, describing prompt patterns and structures, creating new prompts, evaluating existing prompts for improvements, determining if current logic should be extracted to prompts, identifying outdated prompt techniques, optimizing context usage, implementing guardrails, or when user mentions "prompt engineering", "hallucinations", "context optimization", "prompt caching", "chain-of-thought", or asks if logic should become a prompt.4---5
6# Managing Prompts
7
8Comprehensive prompt engineering for Claude 4.5: creation, analysis, optimization, guardrails, context management.
9
10## When to Use
11
12- User requests creating/analyzing/optimizing prompts
13- User asks if current logic should be extracted to prompt
14- User mentions: hallucinations, consistency issues, context optimization, prompt caching, chain-of-thought, XML tags
15- Context usage high and extraction would help
16- Outdated prompt techniques need Claude 4.5 updates
17
18## Core Workflows
19
20### Workflow 1: Analyzing Existing Prompts
21
22**Quick process:**
231. Read prompt file/content
242. Load quick-reference.md for validation checklist
253. Apply evaluation: structure, techniques, guardrails, context efficiency, Claude 4.5 optimization
264. Categorize issues: critical | major | minor
275. Generate recommendations with priority
286. Output using output-formats.md#analysis-report template
29
30**Detailed evaluation:** Load analysis-patterns.md
31
32**Common issues reference:** quick-reference.md#common-issues
33
34---
35
36### Workflow 2: Creating New Prompts
37
38**Quick process:**
391. **Gather requirements:** objective, success criteria, inputs/outputs, complexity, risk factors
402. **Select architecture:** Load architecture-patterns.md, choose: simple_task | complex_reasoning | agent_workflow | multi_window
413. **Apply techniques:** Load technique-reference.md for indexed lookup by need
424. **Add guardrails:** Based on risk level, load guardrails-implementation.md for patterns
435. **Optimize context:** Load context-optimization.md if >10K tokens
446. **Validate:** Use quick-reference.md#quick-validation-checklist
45
46**Technique order:**
471. Clear instructions (explicit, contextual motivation, success criteria, output format)
482. Examples (3-5 for complex tasks)
493. Chain-of-thought (for reasoning tasks)
504. XML tags (for structure)
515. Role assignment, prefilling, template variables
52
53**Architecture selection:**
54- Simple task: Clear objective, straightforward
55- Complex reasoning: Multi-step analysis, judgment
56- Agent workflow: Tool access, iterative, autonomous
57- Multi-window: >100K tokens, extended work
58
59**Guardrails by risk:**
60- Low: Basic uncertainty permission
61- Medium: Examples + format spec
62- High: Quote-grounding + citations
63- Critical: Full stack (quotes, citations, knowledge restriction, CoT)
64
65**Context optimization:**
66- Cache static content ≥1024 tokens
67- Place 20K+ docs at beginning
68- Remove Claude's existing knowledge
69- Quote-grounding for large docs
70- Split if >3 distinct subtasks
71
72---
73
74### Workflow 3: Optimizing Existing Prompts
75
76**Quick process:**
771. Identify goal: reduce_tokens | improve_quality | increase_consistency | reduce_hallucinations | enable_caching | split_prompt
782. Load optimization-strategies.md for before/after examples
793. Apply transformation
804. Measure improvement
815. Document using output-formats.md#optimization-report
82
83**Quick optimization strategies:**
84- **Context reduction:** Remove redundancy, cache static, quote-grounding, split if >3 subtasks
85- **Quality:** Add examples, CoT, explicit instructions, contextual motivation
86- **Consistency:** Format spec, prefilling, 3-5 examples
87- **Hallucinations:** Uncertainty permission, quote-grounding, citations
88- **Caching:** Identify ≥1024 static tokens, restructure, add cache_control
89
90**Decision aids:** quick-reference.md#fast-decision-trees
91
92---
93
94### Workflow 4: Updating Outdated Prompts
95
96**Quick process:**
971. Load migration-guide.md for Claude 3→4.5 patterns
982. Identify outdated: insufficient explicitness, missing motivation, suggestion framing, no summary requests
993. Apply updates preserving core logic
1004. Test and document
101
102**Common migrations:**
103- Vague → Explicit instructions with specifics
104- No motivation → Add contextual reasoning
105- Suggestion framing → Action framing ("Improve this" not "Can you improve?")
106- No summaries → Explicit requests (4.5 less verbose by default)
107- No quality modifiers → Add "Give it your all" for creative tasks
108
109---
110
111### Workflow 5: Deciding Logic Extraction
112
113**Quick process:**
1141. Load extraction-decision-guide.md#decision-tree
1152. Evaluate criteria (6 for prompt, 6 for script)
1163. Score: 3+ = extract, <3 = keep inline
1174. Output using output-formats.md#extraction-decision template
118
119**Prompt extraction criteria (3+ to extract):**
120- Reused across 3+ contexts
121- Needs prompt techniques (examples, CoT, XML)
122- Benefits from guardrails
123- Context optimization beneficial
124- Standardization needed
125- Complex (>200 tokens instructions)
126
127**Script extraction criteria (3+ to extract):**
128- Deterministic/algorithmic
129- Computation-heavy
130- Unit testable
131- Reused across prompts
132- No reasoning required
133- Format conversion/validation
134
135**Keep inline:** One-time, context-specific, simple, requires real-time interaction
136
137## Fast Reference System
138
139**For validation:** quick-reference.md (quick checklist, common issues, decision trees, severity guide)
140
141**For technique lookup:** technique-reference.md (indexed by use case, lookup table, application patterns, combinations)
142
143**For output templates:** output-formats.md (analysis, optimization, extraction, creation, migration reports)
144
145**For detailed patterns:**
146- extended-thinking-implementation.md (internal reasoning, budget management, batch processing, multi-round strategies)
147- prompt-chaining-architecture.md (sequential/parallel workflows, handoff patterns, orchestration, self-correction loops)
148- consistency-techniques.md (prefilling, Structured Outputs, format enforcement, advanced patterns)
149- claude-4-5-optimization.md (4.5-specific guidance: explicit instructions, context awareness, multi-context workflows, tool usage)
150- architecture-patterns.md (5 pattern types: simple task, complex reasoning, agent, multi-window, multi-context workflows)
151- guardrails-implementation.md (hallucination, consistency, security, character patterns, jailbreak prevention, prompt leak reduction)
152- context-optimization.md (token management, caching with model-specific minimums, document org, splitting, cache invalidation patterns)
153- optimization-strategies.md (11 before/after examples)
154- migration-guide.md (Claude 3→4.5 patterns, 10 migration examples)
155- extraction-decision-guide.md (decision trees, criteria, examples)
156- analysis-patterns.md (common issues, quality assessment, evaluation examples)
157
158## Claude 4.5 Critical Behaviors
159
160**Strengths:** Long-horizon reasoning, context awareness, parallel execution, subagent delegation
161
162**Requirements:** More explicit than Claude 3, contextual motivation, high-quality examples, action framing, explicit summary requests, quality modifiers
163
164**Optimization:** Parallel tool guidance, clear success criteria, investigation before answering, aesthetic direction (UI), algorithm approach (code)
165
166## Quick Troubleshooting
167
168Load quick-reference.md for detailed troubleshooting with fixes.
169
170**Common symptoms → Quick fixes:**
171- Inconsistent outputs → Format spec + 3-5 examples
172- Hallucinations → Quote-grounding + uncertainty permission
173- High tokens → Cache static content, remove redundancy
174- Too concise (4.5) → Explicit detail requests
175- Suggests not implements (4.5) → Action framing
176- Stops early (4.5) → Reference token budget available
177
178## Tool Usage Pattern
179
180**Standard workflow:**
1811. Identify which workflow (1-5)
1822. Load quick-reference.md for checklist/decisions
1833. Load technique-reference.md if need technique lookup
1844. Load detailed supporting file only when needed
1855. Use output-formats.md template for deliverable
186
187**Example - Analysis:**
1881. Read prompt
1892. quick-reference.md → validation checklist
1903. technique-reference.md → identify techniques used
1914. analysis-patterns.md → detailed framework (if complex)
1925. output-formats.md#analysis-report → generate output
193
194**Example - Creation:**
1951. Gather requirements
1962. technique-reference.md → select techniques by need
1973. architecture-patterns.md → get template for pattern type
1984. quick-reference.md → validation before finalizing
1995. output-formats.md#creation-summary → document (optional)
200
201**Example - Optimization:**
2021. Identify goal
2032. quick-reference.md#fast-decision-trees → decide strategy
2043. optimization-strategies.md → get before/after pattern
2054. Apply and test
2065. output-formats.md#optimization-report → document
207
208## Validation Before Deployment
209
210Load quick-reference.md#quick-validation-checklist
211
212**Essential checks:**
213- [ ] Objective clear, success criteria defined
214- [ ] Instructions explicit (action framing for 4.5)
215- [ ] Output format specified
216- [ ] Examples if complex (3-5, diverse, identical structure)
217- [ ] CoT if reasoning required
218- [ ] XML if multiple components
219- [ ] Guardrails match risk level
220- [ ] Context optimized (caching, placement, splitting)
221- [ ] Claude 4.5 specificity (explicit, motivation, quality modifiers)
222
223**Golden test:** "Would colleague understand with minimal context?"
224
225## Summary
226
227This 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.
majiayu000