Skill v1.0.0

currentAutomated scan100/100

trinhhaox/ag-kit/context-engineering

──Details

PublishedOctober 13, 2026 at 09:24 AM

Content Hashsha256:9bf922c9f26e5cc3...

Git SHA67b559b415f1

──Files

Files (1 file, 4.8 KB)

SKILL.md4.8 KBactive

SKILL.md · 160 lines · 4.8 KB

version: "1.0.0" name: context-engineering description: LLM context window optimization - budget allocation, compression, progressive loading, RAG retrieval patterns for efficient AI interactions. allowed-tools: Read, Write, Edit, Glob, Grep

Context Engineering

Optimize how information is fed to LLMs for maximum output quality.

Context Budget Model

5-Zone Allocation

┌─────────────────────────────┐
│ System Instructions  10-15% │ Role, rules, output format
├─────────────────────────────┤
│ Task Context         20-30% │ Current request, conversation history
├─────────────────────────────┤
│ Reference Material   30-40% │ Code, docs, examples, search results
├─────────────────────────────┤
│ Working Memory       15-25% │ Intermediate reasoning, tool results
├─────────────────────────────┤
│ Output Reserve       10-15% │ Space for the response
└─────────────────────────────┘

Token Budgeting

typescript

const CONTEXT_LIMIT = 200000; // Claude's context window
const budget = {
  system: Math.floor(CONTEXT_LIMIT * 0.12),
  task: Math.floor(CONTEXT_LIMIT * 0.25),
  reference: Math.floor(CONTEXT_LIMIT * 0.35),
  working: Math.floor(CONTEXT_LIMIT * 0.18),
  output: Math.floor(CONTEXT_LIMIT * 0.10),
};

Compression Techniques

Code Compression

Full file (expensive):
  → Remove comments and blank lines
  → Collapse import blocks to summary
  → Show only changed functions in full
  → Summarize unchanged functions as signatures only

Document Compression

Full document:
  → Extract headings + first sentence per section
  → Keep code examples, remove prose
  → Progressive: summary → relevant sections → full text

Conversation Compression

Turn 1-5 (old): Summarize to 2-3 sentences
Turn 6-8 (recent): Keep key decisions and outcomes
Turn 9-10 (current): Keep verbatim

Progressive Loading Strategy

Level 1: Structural Overview

Read file tree → understand project layout
Read function signatures → understand API surface
Read type definitions → understand data model

Level 2: Targeted Detail

Read specific function bodies relevant to task
Read test files for the module being modified
Read config files affecting behavior

Level 3: Deep Context

Read related modules for side effects
Read git history for recent changes
Read documentation for design decisions

Decision Rule

Load the minimum context needed to make a correct decision.
If uncertain, load one more level, not everything.

RAG (Retrieval-Augmented Generation)

Retrieval Strategy (Priority Order)

Exact match — function name, error message, file path
Semantic search — vector embedding similarity
Keyword search — BM25 / full-text search
Structural search — AST-based code search

Document Ranking

typescript

type SearchResult = {
  content: string;
  score: number;       // Relevance score
  source: string;      // File path or URL
  freshness: number;   // How recent (0-1)
  authority: number;   // Source trustworthiness (0-1)
};
// Composite ranking
const finalScore = (r: SearchResult) =>
  r.score * 0.5 + r.freshness * 0.3 + r.authority * 0.2;

Chunking Strategies

Strategy	Chunk Size	Overlap	Use When
Fixed-size	500-1000 tokens	100 tokens	General text
Semantic	By paragraph/section	None	Structured docs
Code-aware	By function/class	Include imports	Source code
Recursive	Split by headers, then paragraphs	Minimal	Long documents

Context Quality Signals

Include (High Value)

Type definitions and interfaces
Function signatures with JSDoc
Error messages and stack traces
Test cases (show expected behavior)
Recent git diff (shows intent)

Exclude (Low Value)

Lock files (package-lock.json)
Generated code (dist/, .next/)
Binary files
Redundant imports across files
Unchanged boilerplate

Anti-Patterns

Anti-Pattern	Problem	Fix
Dump entire file	Wastes context, dilutes signal	Extract relevant functions only
No context	Hallucination, incorrect assumptions	Provide minimum viable context
Stale context	Outdated info causes wrong decisions	Refresh before each major step
Repeated context	Same info in system + task + reference	Deduplicate across zones
Over-retrieval	Too many RAG results overwhelm	Top-k=5, re-rank, filter by score threshold

All versions