context-engineering

Name: context-engineering
Author: itsmostafa

25installs0forks

How to install context-engineering?

itsmostafa/llm-engineering-skills/context-engineering

$npx -y skills add itsmostafa/llm-engineering-skills --skill context-engineering

Installs into the current project.

Summary

Strategies for managing LLM context windows effectively in AI agents. Use when building agents that handle long conversations, multi-step tasks, tool orchestration, or need to maintain coherence across extended interactions.

Files · 1

View on GitHub

SKILL.md

1	# Context Engineering
2
3	Context engineering is the discipline of curating and maintaining the optimal set of tokens during LLM inference. Unlike prompt engineering (crafting individual prompts), context engineering focuses on what information enters the context window and when.
4
5	## Table of Contents
6
7	- [Core Principles](#core-principles)
8	- [Context Management Strategies](#context-management-strategies)
9	- [System Prompt Design](#system-prompt-design)
10	- [Tool Design for Context Efficiency](#tool-design-for-context-efficiency)
11	- [Long-Horizon Task Patterns](#long-horizon-task-patterns)
12	- [Implementation Patterns](#implementation-patterns)
13	- [Best Practices](#best-practices)
14	- [References](#references)
15
16	## Core Principles
17
18	### Context as a Finite Resource
19
20	LLMs have limited "attention budgets." As context length increases, models experience context rot—decreased ability to accurately recall information. The goal is finding the smallest possible set of high-signal tokens that maximize desired outcomes.
21
22	```
23	Effective Context = Relevant Information / Total Tokens
24	```
25
26	Key insight: More context isn't better. The right context is better.
27
28	### The Context Pollution Problem
29
30	Every token added to context has costs:
31	- Increased latency and compute
32	- Diluted attention to important information
33	- Higher risk of hallucination from conflicting data
34	- Reduced model performance on retrieval tasks
35
36	## Context Management Strategies
37
38	### 1. Context Trimming
39
40	Drop older conversation turns, keeping only the last N turns.
41
42	\| Aspect \| Details \|
43	\|--------\|---------\|
44	\| Mechanism \| Sliding window over conversation history \|
45	\| Pros \| Deterministic, zero latency, preserves recent context verbatim \|
46	\| Cons \| Abrupt loss of long-range context, "amnesia" effect \|
47	\| Best for \| Independent tasks, short interactions, predictable workflows \|
48
49	```python
50	def trim_context(messages: list, keep_last_n: int = 10) -> list:
51	"""Keep system message + last N turns."""
52	system_msgs = [m for m in messages if m["role"] == "system"]
53	other_msgs = [m for m in messages if m["role"] != "system"]
54	return system_msgs + other_msgs[-keep_last_n:]
55	```
56
57	### 2. Context Summarization
58
59	Compress prior messages into structured summaries.
60
61	\| Aspect \| Details \|
62	\|--------\|---------\|
63	\| Mechanism \| LLM generates summary of older context \|
64	\| Pros \| Retains long-range memory, smoother UX, scalable \|
65	\| Cons \| Summarization bias risk, added latency, potential compounding errors \|
66	\| Best for \| Complex multi-step tasks, long-horizon interactions \|
67
68	```python
69	SUMMARIZATION_PROMPT = """Summarize the conversation so far, preserving:
70	1. Key decisions made
71	2. Important context established
72	3. Current task state and goals
73	4. Any constraints or preferences expressed
74
75	Be concise but complete. Output as structured markdown."""
76
77	async def summarize_context(messages: list, model) -> str:
78	"""Generate a summary of conversation history."""
79	conversation_text = format_messages_for_summary(messages)
80	response = await model.generate(
81	system=SUMMARIZATION_PROMPT,
82	user=conversation_text
83	)
84	return response.content
85	```
86
87	### 3. Hybrid Approach
88
89	Combine trimming and summarization for optimal balance.
90
91	```python
92	class HybridContextManager:
93	def __init__(
94	self,
95	keep_recent: int = 5, # Recent turns to keep verbatim
96	summary_threshold: int = 20, # When to trigger summarization
97	):
98	self.keep_recent = keep_recent
99	self.summary_threshold = summary_threshold
100	self.running_summary = ""
101
102	def process(self, messages: list) -> list:
103	if len(messages) < self.summary_threshold:
104	return messages
105
106	# Summarize older messages
107	old_messages = messages[:-self.keep_recent]
108	self.running_summary = summarize(old_messages, self.running_summary)
109
110	# Return summary + recent messages
111	return [
112	{"role": "system", "content": f"Previous context:\n{self.running_summary}"},
113	*messages[-self.keep_recent:]
114	]
115	```
116
117	### 4. Session Memory
118
119	Persist reusable facts, preferences, and task state outside the context window. Load only the relevant slice for the current turn.
120
121	\| Aspect \| Details \|
122	\|--------\|---------\|
123	\| Mechanism \| External store keyed by user, session, task, or resource \|
124	\| Pros \| Recovers long-range context without carrying all history \|
125	\| Cons \| Requires retrieval, freshness, and deletion policies \|
126	\| Best for \| Agents, project work, personalization, long-running workflows \|
127
128	Separate durable memory from ephemeral scratchpads. Durable memory should contain stable facts and explicit decisions, not every intermediate thought.

People who liked this also viewed

6 picks

Type

context-engineering

25installs0forks

View on GitHub

How to install context-engineering?

itsmostafa/llm-engineering-skills/context-engineering

$npx -y skills add itsmostafa/llm-engineering-skills --skill context-engineering

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Context Engineering
2
3	Context engineering is the discipline of curating and maintaining the optimal set of tokens during LLM inference. Unlike prompt engineering (crafting individual prompts), context engineering focuses on what information enters the context window and when.
4
5	## Table of Contents
6
7	- [Core Principles](#core-principles)
8	- [Context Management Strategies](#context-management-strategies)
9	- [System Prompt Design](#system-prompt-design)
10	- [Tool Design for Context Efficiency](#tool-design-for-context-efficiency)
11	- [Long-Horizon Task Patterns](#long-horizon-task-patterns)
12	- [Implementation Patterns](#implementation-patterns)
13	- [Best Practices](#best-practices)
14	- [References](#references)
15
16	## Core Principles
17
18	### Context as a Finite Resource
19
20	LLMs have limited "attention budgets." As context length increases, models experience context rot—decreased ability to accurately recall information. The goal is finding the smallest possible set of high-signal tokens that maximize desired outcomes.
21
22	```
23	Effective Context = Relevant Information / Total Tokens
24	```
25
26	Key insight: More context isn't better. The right context is better.
27
28	### The Context Pollution Problem
29
30	Every token added to context has costs:
31	- Increased latency and compute
32	- Diluted attention to important information
33	- Higher risk of hallucination from conflicting data
34	- Reduced model performance on retrieval tasks
35
36	## Context Management Strategies
37
38	### 1. Context Trimming
39
40	Drop older conversation turns, keeping only the last N turns.
41
42	\| Aspect \| Details \|
43	\|--------\|---------\|
44	\| Mechanism \| Sliding window over conversation history \|
45	\| Pros \| Deterministic, zero latency, preserves recent context verbatim \|
46	\| Cons \| Abrupt loss of long-range context, "amnesia" effect \|
47	\| Best for \| Independent tasks, short interactions, predictable workflows \|
48
49	```python
50	def trim_context(messages: list, keep_last_n: int = 10) -> list:
51	"""Keep system message + last N turns."""
52	system_msgs = [m for m in messages if m["role"] == "system"]
53	other_msgs = [m for m in messages if m["role"] != "system"]
54	return system_msgs + other_msgs[-keep_last_n:]
55	```
56
57	### 2. Context Summarization
58
59	Compress prior messages into structured summaries.
60
61	\| Aspect \| Details \|
62	\|--------\|---------\|
63	\| Mechanism \| LLM generates summary of older context \|
64	\| Pros \| Retains long-range memory, smoother UX, scalable \|
65	\| Cons \| Summarization bias risk, added latency, potential compounding errors \|
66	\| Best for \| Complex multi-step tasks, long-horizon interactions \|
67
68	```python
69	SUMMARIZATION_PROMPT = """Summarize the conversation so far, preserving:
70	1. Key decisions made
71	2. Important context established
72	3. Current task state and goals
73	4. Any constraints or preferences expressed
74
75	Be concise but complete. Output as structured markdown."""
76
77	async def summarize_context(messages: list, model) -> str:
78	"""Generate a summary of conversation history."""
79	conversation_text = format_messages_for_summary(messages)
80	response = await model.generate(
81	system=SUMMARIZATION_PROMPT,
82	user=conversation_text
83	)
84	return response.content
85	```
86
87	### 3. Hybrid Approach
88
89	Combine trimming and summarization for optimal balance.
90
91	```python
92	class HybridContextManager:
93	def __init__(
94	self,
95	keep_recent: int = 5, # Recent turns to keep verbatim
96	summary_threshold: int = 20, # When to trigger summarization
97	):
98	self.keep_recent = keep_recent
99	self.summary_threshold = summary_threshold
100	self.running_summary = ""
101
102	def process(self, messages: list) -> list:
103	if len(messages) < self.summary_threshold:
104	return messages
105
106	# Summarize older messages
107	old_messages = messages[:-self.keep_recent]
108	self.running_summary = summarize(old_messages, self.running_summary)
109
110	# Return summary + recent messages
111	return [
112	{"role": "system", "content": f"Previous context:\n{self.running_summary}"},
113	*messages[-self.keep_recent:]
114	]
115	```
116
117	### 4. Session Memory
118
119	Persist reusable facts, preferences, and task state outside the context window. Load only the relevant slice for the current turn.
120
121	\| Aspect \| Details \|
122	\|--------\|---------\|
123	\| Mechanism \| External store keyed by user, session, task, or resource \|
124	\| Pros \| Recovers long-range context without carrying all history \|
125	\| Cons \| Requires retrieval, freshness, and deletion policies \|
126	\| Best for \| Agents, project work, personalization, long-running workflows \|
127
128	Separate durable memory from ephemeral scratchpads. Durable memory should contain stable facts and explicit decisions, not every intermediate thought.

People who liked this also viewed

6 picks

Type