???? Core Principles Context Window Limited token budget (~200k tokens typical) Context = Input + Output combined Earlier tokens may be "forgotten" (lost attention) Most recent context has…

???? Core Principles

Context Window

Limited token budget (~200k tokens typical)
Context = Input + Output combined
Earlier tokens may be “forgotten” (lost attention)
Most recent context has highest weight

Information Density

Dense, structured input > verbose
Use XML/JSON for complex data
Key info at start AND end
Remove redundancy ruthlessly

???? Prompt Structure

Optimal Template

<role>Expert persona</role> <task>What to do</task> <context>Background info</context> <constraints>Rules</constraints> <output_format>Structure</output_format> <examples>…</examples>

Critical Elements

Clear objective (single, specific)
Success criteria defined
Format requirements explicit
Edge cases addressed

???? Few-Shot Techniques

Example Types

Positive: Ideal outputs
Negative: What to avoid
Edge cases: Boundary conditions
Progressive: Simple → Complex

Best Practices

3-5 examples optimal
Diverse, representative samples
Annotate reasoning process
Match target complexity

???? Chain-of-Thought

Activation Phrases

Zero-shot: “Let’s think step by step…”
“First, analyze… Then…”
“Break this down systematically”

Advanced Patterns

Self-Consistency: Multiple reasoning paths
Tree-of-Thought: Explore branches
Meta-prompting: Analyze prompt first

⚙️ Token Optimization

Compression Strategies

Use abbreviations consistently
Remove articles (a, an, the)
Prefer numerals over words
Use symbols: → ∴ ∵ ≈

Token Count Rules

~4 chars = 1 token
100 words ≈ 75 tokens

Whitespace counts
Code is token-expensive
Repeated patterns compress

???? Role & Persona Engineering

Effective Personas

Generic: “You are helpful”
Specific: “You are a senior DevOps engineer with 15 years experience in Kubernetes, known for writing clear, actionable documentation”

Specific expertise domain
Years of experience
Communication style
Known strengths/weaknesses

Multi-Agent Patterns

Debate: Pro vs Con roles
Specialist Panel: Expert consultation
Red Team: Adversarial review
Socratic: Question-driven dialogue

Analyst: {analysis} Critic: {critique} Synthesizer: {final_output}

???? Output Control

Format Specifications

Structured: JSON, XML, YAML
Length: Exact word/token count
Style: Tone, formality, POV
Constraints: No speculation, cite sources

Validation Triggers

“Before responding, verify:
✓ All claims sourced
✓ Format matches schema
✓ Edge cases handled”

????️ Context Window Management

Prioritization Strategies

Recency: Latest messages matter most
Relevance: Semantic similarity search
Importance: Tag critical information
Summarization: Compress old context

Priority_Score = (Recency × 0.3) + (Relevance × 0.5) + (Importance × 0.2)

Memory Patterns

Sliding Window: Keep N recent turns
Hierarchical: Summary + recent detail
Retrieval: Vector DB + current context
Compression: Distill multi-turn to summary

System: {permanent_instructions} Summary: {condensed_history} Recent: {last_3_turns} Current: {user_message}

???? RAG (Retrieval)

Context Injection

Embed retrieved docs with metadata
Source attribution required
Chunk size: 200-500 tokens optimal
Include relevance scores

Quality Signals

<source confidence=”0.89″>
<metadata>Updated: 2024-01</metadata>
{retrieved_content}
</source>

???? Advanced Methods

Self-Reflection

1. Generate initial response 2. Critique: “What’s wrong?” 3. Refine based on critique 4. Repeat if needed

Constitutional AI

Embed principles in system
Self-critique against rules
Revision loops

???? Semantic Control

Embedding Strategies

Similar examples cluster together
Use for dynamic few-shot selection
Semantic search > keyword match
Rerank by similarity score

Distance Metrics

Cosine Similarity Euclidean Dot Product

⚠️ Common Pitfalls & Solutions

❌ Vague Instructions

Bad: “Write something good”
Good: “Write a 500-word technical blog post for developers explaining OAuth 2.0 flow with code examples in Python”

❌ Conflicting Directives

Bad: “Be brief but comprehensive”
Good: “Provide a 200-word summary covering the 3 most important points”

❌ Assumed Knowledge

Bad: “Fix the bug”
Good: “The login form submits but returns 401. Here’s the code: {…}”

❌ No Format Spec

Always define output structure
Use schema/template
Example of desired format

❌ Ignoring Tokens

Monitor context usage
Compress when possible
Truncate intelligently

❌ One-Shot Hoping

Iterate prompts
Test variations
A/B test approaches

???? Testing & Evaluation

Evaluation Metrics

Accuracy: Factual correctness
Relevance: On-topic responses
Coherence: Logical flow
Consistency: Same input → same output
Completeness: All requirements met

Testing Strategies

Build test suite of edge cases
Use golden datasets
Temperature testing (0.0 → 1.0)
Automated regression tests
Human-in-loop validation

⚡ Quick Wins

Instant Improvements

✓ Add “Let’s think step-by-step”
✓ Provide 2-3 examples
✓ Define output format
✓ Use XML tags for structure
✓ Specify desired length
✓ Include validation criteria

???? Model-Specific Strategies

Claude-Specific

Prefers XML structure
Excels at long context
Strong with technical tasks
Use <thinking> tags for reasoning

GPT-Specific

Follows system messages strongly
JSON mode available
Function calling native
Sensitive to instruction order

????️ Parameters

Temperature Guide

0.0-0.3: Deterministic, factual
0.4-0.7: Balanced creativity
0.8-1.0: Highly creative

Other Parameters

top_p: Nucleus sampling (0.9 typical)
max_tokens: Hard output limit
frequency_penalty: Reduce repetition

???? Security & Safety

Prompt Injection Defense

Sandwich user input with instructions
Use delimiters clearly
Validate output format strictly
Never trust user-provided “system” messages

Content Filtering

<user_input> {sanitized_input} </user_input> Respond only to legitimate requests

AI Context Engineering – Ultimate Cheat Sheet