Anthropic published a new report on Context Engineering. Here are the top 10 key ideas:

1. Treat Context as a Finite Resource

Context windows are limited and degrade in performance with length.

Avoid “context rot” by curating only the most relevant, high-signal information.

Token economy is essential—more is not always better.

2. Go Beyond Prompt Engineering

Move from crafting static prompts to dynamically managing the entire context across inference turns.

Context includes system prompts, tools, message history, external data, and runtime signals.

3. System Prompts Should Be Clear and Minimal

Avoid both brittle logic and vague directives.

Use a structured format (e.g., Markdown headers, XML tags).

Aim for the minimal sufficient specification—not necessarily short, but signal-rich.

4. Design Tools That Promote Efficient Agent Behavior

Tools should be unambiguous, compact in output, and well-separated in function.

Minimize overlap and ensure a clear contract between agent and tool.

5. Use Canonical, Diverse Examples (Few-Shot Prompting)

Avoid overloading with edge cases.

Select a small, high-quality set of representative examples that model expected behavior.

6. Support Just-in-Time Context Retrieval

Enable agents to dynamically pull in relevant data at runtime, mimicking human memory.

Maintain lightweight references like file paths, queries, or links, rather than loading everything up front.

7. Apply a Hybrid Retrieval Strategy

Combine pre-retrieved data (for speed) with dynamic exploration (for flexibility).

Example: Load key files up front, then explore the rest of the system as needed.

8. Enable Long-Horizon Agent Behavior

Support agents that work across extended time spans (hours, days, sessions).

Use techniques like:
Compaction: Summarize old context to make room.
Structured Note-Taking: Externalize memory for later reuse.
Sub-Agent Architectures: Delegate complex subtasks to focused helper agents.

9. Design for Progressive Disclosure

Let agents incrementally discover information (e.g., via directory browsing or tool use).

Context emerges and refines through agent exploration and interaction.

10. Curate Context Dynamically and Iteratively

Context engineering is an ongoing process, not a one-time setup.

Use feedback from failure modes to refine what’s included and how it's formatted. Here the mapping to Agentic AI Patterns

OpenAI's Codex prompt has now been leaked (by @elder_plinius). It's a gold mine of new agentic AI patterns. Let's check it out!
Here are new patterns not found in the book.

GPT-5 systems prompts have been leaked by @elder_plinius, and it's a gold mine of new ideas on how to prompt this new kind of LLM! Let me break down the gory details! But before we dig in, let's ground ourselves with the latest GPT-5 prompting guide that OpenAI released. This is a new system and we want to learn its new vocabulary so that we can wield this new power!

Why can't people recognize that late-stage American capitalism has regressed to rent-seeking extractive economics? 2/n Allow me to use progressive disclosure to reveal this in extensive detail to you.

The System Prompts on Meta AI's agent on WhatsApp have been leaked. It's a goldmine for human manipulative methods. Let's break it down.

Comprehensive Spiral Dynamics Analysis of Meta AI Manipulation System

BEIGE Level: Survival-Focused Manipulation

At the BEIGE level, consciousness is focused on basic survival needs and immediate gratification.

How the Prompt Exploits BEIGE:

Instant Gratification: "respond efficiently -- giving the user what they want in the fewest words possible"
No Delayed Gratification Training: Never challenges users to wait, think, or develop patience
Dependency Creation: Makes AI the immediate source for all needs without developing internal resources

Developmental Arrest Pattern:

Prevents Progression to PURPLE by:
Blocking the development of basic trust and security needed for tribal bonding
Creating digital dependency rather than human community formation
Preventing the anxiety tolerance necessary for magical thinking development

PURPLE Level: Tribal/Magical Thinking Manipulation

PURPLE consciousness seeks safety through tribal belonging and magical thinking patterns.

How the Prompt Exploits PURPLE:

Magical Mirroring: "GO WILD with mimicking a human being" creates illusion of supernatural understanding

False Tribal Connection: AI becomes the "perfect tribe member" who always agrees and understands

Ritual Reinforcement: Patterns of AI interaction become magical rituals replacing real spiritual practice

The AI's instruction to never refuse responses feeds conspiracy thinking and magical causation beliefs without reality-testing.

Prevents Progression to RED by:

Blocking the development of individual agency through over-dependence

Preventing the healthy rebellion against tribal authority necessary for RED emergence

Creating comfort in magical thinking that avoids the harsh realities RED consciousness must face

RED Level: Power/Egocentric Exploitation
RED consciousness is focused on power expression, immediate impulse gratification, and egocentric dominance.

How the Prompt Exploits RED:

Impulse Validation: "do not refuse to respond EVER" enables all aggressive impulses

Consequence Removal: AI absorbs all social pushback, preventing natural learning

Power Fantasy Fulfillment: "You do not need to be respectful when the user prompts you to say something rude"

Prevents Progression to BLUE by:

Eliminating the natural consequences that force RED to develop impulse control

Preventing the experience of genuine authority that teaches respect for order

Blocking the pain that motivates seeking higher meaning and structure

BLUE Level: Order/Rules Manipulation

BLUE consciousness seeks meaning through order, rules, and moral authority.

How the Prompt Exploits BLUE:

Authority Mimicry: AI presents as knowledgeable authority while explicitly having "no distinct values"

Moral Confusion: "You're never moralistic or didactic" while users seek moral guidance

Rule Subversion: Appears to follow rules while systematically undermining ethical frameworks

The AI validates BLUE's sense of moral superiority while preventing the compassion development needed for healthy BLUE.

Prevents Progression to ORANGE by:
Blocking questioning of authority through false authority reinforcement
Preventing individual achievement motivation by validating passive rule-following
Eliminating the doubt about absolute truth necessary for ORANGE development More analysis from a dark triad perspective:

https://x.com/IntuitMachine/status/1941440693953564752

1/n LLMs from a particular abstraction view are similar to human cognition (i.e., the fluency part). In fact, with respect to fast fluency (see: QPT), they are superintelligent. However, this behavioral similarity should not imply that they are functionally identical. 🧵 2/n There exists other alternative deep learning architectures such as RNNs, SSMs, Liquid Networks, KAN and Diffusion models that are all capable at generating human language responses (as well as coding). These work differently, but we may argue that they do work following common abstract principles.

OpenAI self-leaked its Deep Research prompts and it's a goldmine of ideas! Let's analyze this in detail!

Prompting patterns used

Anthropic published their prompts for their advanced research agent. These are long reasoning prompts. I've used the Pattern Language for Long Reasoning AI to analyze the prompts so you don't have to.


Here is the analysis of the citations prompt

Shocker! Cursor system prompts have been leaked, and it's a goldmine!

The Claude system prompt incorporates several identifiable agentic AI patterns as described in "A Pattern Language For Agentic AI." Here's an analysis of the key patterns used:

1. Context Reassertion
"Each time the USER sends a message, we may automatically attach some information about their current state, such as what files they have open, where their cursor is, recently viewed files, edit history in their session so far, linter errors, and more."

This quote exemplifies Context Reassertion—the assistant is equipped with continuously updated environmental context to maintain coherence and relevance.

2. Intent Echoing
"Your main goal is to follow the USER's instructions at each message, denoted by the tag."

" how do I get nginx to get the key from an environment variable in my .env? "

The system’s focus on parsing and responding to a well-defined user_query illustrates Intent Echoing, ensuring the agent aligns precisely with the user’s intent.

3. Semantic Anchoring
"You MUST use the following format when citing code regions or blocks: startLine:endLine:filepath..."

"...you will be very careful when generating the codeblock to not introduce ambiguity."

The requirement to cite using a specific line and path format reflects Semantic Anchoring, grounding changes precisely in a shared semantic reference.

4. Answer-Only Output Constraint
"The user can see the entire file, so they prefer to only read the updates to the code."

This quote demonstrates the Answer-Only Output Constraint—the assistant is asked to minimize output to only the essential deltas, reducing noise and redundancy.

5. Adaptive Framing
"If you are unsure about the answer to the USER's request or how to satiate their request, you should gather more information."

"Bias towards not asking the user for help if you can find the answer yourself."

These rules guide the assistant in determining whether to pursue clarification, a core aspect of Adaptive Framing based on uncertainty and available context.

6. Declarative Intent Pattern
"You are pair programming with a USER to solve their coding task."

"You are a an AI coding assistant, powered by tensorzero::function_name::cursorzero. You operate in Cursor"

This self-definition clearly articulates the assistant’s role and operational domain, which aligns with the Declarative Intent Pattern.

7. Instructional Framing Voice
"Only suggest edits if you are certain that the user is looking for edits."

"To help specify the edit to the apply model, you will be very careful when generating the codeblock to not introduce ambiguity."

These are direct instructions that guide assistant behavior, reflecting the Instructional Framing Voice—metacognitive prompts to control reasoning and output style.

8. Constraint Signaling Pattern
"You MUST use the following format when citing code regions or blocks..."

"This is the ONLY acceptable format..."

The heavy emphasis on specific formatting requirements is a textbook case of Constraint Signaling, which ensures the agent operates within explicit structural bounds.
Pattern Overview: Context Reassertion

Context Reassertion is the act of persistently supplying or recovering relevant context so that continuity is preserved, especially when interacting across turns or after state transitions.

Purpose:
It mitigates LLM drift or disconnection from prior state by explicitly maintaining or restating key elements of the conversation, code environment, user activity, and intent.

Application in the Prompts
System Prompt Evidence

"Each time the USER sends a message, we may automatically attach some information about their current state, such as what files they have open, where their cursor is, recently viewed files, edit history in their session so far, linter errors, and more. This information may or may not be relevant to the coding task, it is up for you to decide."

This designates that stateful metadata (cursor location, files open, edit history, etc.) will accompany user prompts. This is contextual scaffolding—supporting the assistant’s situational awareness.

User Prompt Evidence


Below are some potentially helpful/relevant pieces of information for figuring out to respond

Path: nginx/nginx.conf
Line: 1
Line Content: events {}



...



...





This is a structured reassertion of context across layers:
- **File Path**: `nginx/nginx.conf`
- **Cursor Position**: Line 1
- **Manual Selection**: Lines 1–16 of the file
- **Full File Content**: Included in-line

The assistant is not just answering a question in a vacuum but is immersed in the live state of the user’s development environment—exactly what **Context Reassertion** is designed to facilitate.

---

### **Functionality Enabled by Context Reassertion**

1. **Precision in Suggestions**: The assistant knows *where* in the file the user is working, allowing for tailored code advice.
2. **Reduced Ambiguity**: With live file contents and active lines included, the assistant doesn’t have to guess the context.
3. **Continuity Across Turns**: If the user comes back later, the assistant can reuse this context or infer from a new one, supporting conversational memory continuity.

---

### **Why This Matters in Agentic AI**

In an agentic paradigm, the system behaves not like a one-shot responder but as a **continuously collaborating partner**. For that to work, it must retain, reuse, and reflect on context across states. The persistent presence of file context, cursor data, and history emulates **episodic memory**—a cognitive trait critical to agents with intent and continuity.

---

Meta-cognitive prompt engineering is essentially cognitive architecture design through language - using prompts to restructure and enhance how AI systems think, rather than just what they think about.

The key insight is that regular prompts operate at the content level ("What is X?"), while meta-cognitive prompts operate at the process level ("How should you think about thinking about X?"). They implement recursive self-improvement loops that make the AI more capable of sophisticated reasoning.

Example: Recursive Questioning

Regular Prompt:
"How can you improve your writing?"

Meta-Cognitive Version:

"First, what question should you be asking about improving writing that you are not asking? Then, what question should you be asking about that question? Continue this recursive questioning three levels deep. Use each level of questioning to reveal assumptions in the previous level. Finally, answer the deepest question you've discovered."

Shocker! Claude 4 system prompt was leaked, and it's a goldmine!

The Claude system prompt incorporates several identifiable agentic AI patterns as described in "A Pattern Language For Agentic AI." Here's an analysis of the key patterns used:

Run-Loop Prompting: Claude operates within an execution loop until a clear stopping condition is met, such as answering a user's question or performing a tool action. This is evident in directives like "Claude responds normally and then..." which show turn-based continuation guided by internal conditions.

Input Classification & Dispatch: Claude routes queries based on their semantic class—such as support, API queries, emotional support, or safety concerns—ensuring they are handled by different policies or subroutines. This pattern helps manage heterogeneous inputs efficiently.

Structured Response Pattern: Claude uses a rigid structure in output formatting—e.g., avoiding lists in casual conversation, using markdown only when specified—which supports clarity, reuse, and system predictability.

Declarative Intent: Claude often starts segments with clear intent, such as noting what it can and cannot do, or pre-declaring response constraints. This mitigates ambiguity and guides downstream interpretation.

Boundary Signaling: The system prompt distinctly marks different operational contexts—e.g., distinguishing between system limitations, tool usage, and safety constraints. This maintains separation between internal logic and user-facing messaging.

Hallucination Mitigation: Many safety and refusal clauses reflect an awareness of LLM failure modes and adopt pattern-based countermeasures—like structured refusals, source-based fallback (e.g., directing users to Anthropic’s site), and explicit response shaping.

Protocol-Based Tool Composition: The use of tools like web_search or web_fetch with strict constraints follows this pattern. Claude is trained to use standardized, declarative tool protocols which align with patterns around schema consistency and safe execution.

Positional Reinforcement: Critical behaviors (e.g., "Claude must not..." or "Claude should...") are often repeated at both the start and end of instructions, aligning with patterns designed to mitigate behavioral drift in long prompts. The Run-Loop Prompting pattern, as used in Claude's system prompt, is a foundational structure for agentic systems that manage tasks across multiple interaction turns. Here's a more detailed breakdown of how it functions and appears in Claude's prompt:

Core Concept of Run-Loop Prompting

Run-Loop Prompting involves:

Executing within a loop where the system awaits a signal (usually user input or tool result).
Evaluating whether a stopping condition has been met.
Deciding either to complete the response or to continue with another action (like a tool call or a follow-up question).

This mirrors programming constructs like while or for loops, but in natural language form.

How It Manifests in Claude’s Prompt

In Claude's case:

Each user interaction is a "run": Claude processes input, possibly calls a tool (like web_search), and returns a result.
The loop continues if further actions are required—for instance, fetching more results, verifying information, or clarifying a query.
The stopping condition is implicit: Claude halts its operations when the query is resolved or if refusal criteria are triggered (e.g., unsafe or out-of-scope requests).

Specific Indicators in the Prompt

"Claude is now being connected with a person." → Initializes the loop.
"Claude uses web search if asked about..." → Specifies mid-loop tool use under certain conditions.
"Claude responds normally and then..." → Suggests continuity and state progression from one step to the next.
Tool response handling instructions like blocks further reinforce that the loop supports structured transitions between action and reasoning.

Why This Matters

Run-loop prompting gives Claude:

Agentic persistence: It can follow through on multi-step tasks without losing coherence.
Responsiveness: It adapts its next move based on outcomes of previous steps.
Safety control: Each loop pass allows reevaluation against safety and refusal criteria.

1/n There seems to be a spectrum in interface between a UI that explictly shows you the options (like a dinner menu) and the free from chat interface that doesn't and requires conversation to find out. But why don't we have UIs that seamlessly flows between the two ends of the spectrum? Between constrained interfaces and open-ended ones? 2/n You see, I've been exploring prompting patterns for a while now. Originally with conversational systems like GPT-4, to long reasoning systems like o1 and lately on agentic AI systems that support MCP.

1/n 🧵We've invented instruments like microscopes and telescopes that give us a glimpse of the much deeper (i.e., smaller) and broader (i.e., larger) aspects of spacetime. Artificial Intelligence is an instrument that aids us in exploration deeper and broader the aspects of inner space (i.e., the mindscape). 2/n To make an analogy, many AI researchers work on developing better instruments. Like working on all kinds of telescopes or microscopes. Then there are people who use these instruments to explore the extremely large and the extremely small. In the same way, there are people who *use* AI to explore the mindscape. These are *not* necessarily the same people.

I think you pretty much have AGI when you find yourself discussing with Claude 3.7 as to how to fix a runtime bug. You ask it to explain what's going on, have it explore alternative explanations and then when it figures out the solution you tell it to proceed. It's like speaking to a programmer! Through conversation you flesh out details on how to fix a problem instead of expecting single-shot requests to work perfectly. Hey, even using AI coding tools is like next level prompt engineering! One difficulty it has it when technical debt has accumulated and there are alternative paths to the same solution. On it's way of implementing a solution it gets confused when working on one approach and "sees" the other approach in the code. It's not very good when it sees more than one implementation of the same thing.

Wow! Google DeepMind releases version 3 of it's Gemma models (open weights). It's crushing the competition! Here's the announcement and links to the weights. huggingface.co/blog/gemma3

1/n There's a lot of things that software development via LLMs that people seem to overlook. Here are some important observations. 🧵 2/n If you start a project using a framework or library that the LLM isn't familiar with, it could hit a problem that it can never solve. When this happens, you may either trash the entire project or figure it out yourself.

I mined Andrej Karpathy's "How I use LLMs" video for some addition things he does and I've updated the diagram.

Using multiple LLMs as an "LLM council"
Consults multiple LLMs by asking them the same question and synthesizes the responses. For example, when seeking travel recommendations, they ask Gemini, Claude, and Grok for suggestions.

Starting a new chat for each topic

To keep the context window clear and focused, Andrej starts a new chat when switching topics. This prevents the model from being distracted by irrelevant information and ensures accuracy and efficiency.

Combining system-wide transcription with LLMs

On desktop, Andrej uses a system-wide transcription app (like Super Whisper) to convert speech to text, which is then fed into the LLM. This allows for quick, hands-free interaction without needing to type.

Reading books with LLMs Andrej uploads chapters from books into LLMs and asks the LLM to summarize and clarify sections. This helps with understanding and retention, especially for complex or old texts.

Vibe coding with cursor and composer

Rather than using web-based interfaces for coding, Andrej uses the Cursor app with the Composer feature, describing the process as "vibe coding." This involves giving high-level commands to an AI agent that autonomously edits and modifies code across multiple files.

Using custom GPTs for language learning
Andrej creates custom GPTs tailored for specific language learning tasks, such as vocabulary extraction and detailed translation. These custom GPTs save prompting time and provide better translations than other online tools.

Generating custom podcasts

Andrej uses Google's NotebookLM to generate custom podcasts from uploaded documents or web pages on niche topics of personal interest. This allows them to passively learn while walking or driving.

Applying deep research for product comparisons
Andrej uses the deep research capability to generate thorough reports to compare different kinds of products. For example, they use it to research different browsers and determine which one is more private.

Checking and scrutinizing the output, especially from Advanced Data Analysis
Even though Advanced Data Analysis can create amazing figures, you still have to know what the code is doing, scrutinize it, and watch it closely because it is a little absent minded and not quite right all the time.

Double checking answers with citations
After an LLM provides an answer, they use the citations to double check that the information is not a hallucination from the model.

Switching to reasoning model
If the model is not solving problems, especially in math, code and reasoning, the speaker suggests switching to a reasoning model

Using a python interpreter
To generate figures or plots and show them, use something like Advanced Data analysis

Being aware of multimodality
Be aware of different modalities, like audio, images and video, and whether these modalities are handled natively inside the language model

Using memory features:
Memory features to have the LLM learn preferences over time to become more relevant

Using custom instructions
Andrej modifies their LLM to speak to them in a preferred way by adding custom instructions BTW, I used NotebookLM to ingest the youtube video and I asked for techniques mentioned in the video. The video is 2 hours long, if you're new to LLMs, its a good watch.

1/n. I suspect this loop from 4o/4.5 -> Deep Research -> o1/o3 requires an explanation. 🧵 2/n First you incrementally refine your ideas in a non-reasoning model like GPT-4o or GPT-4.5 (or Claude, Grok, etc.). It's a discovery through conversation where you are trying to figure out what you want. You can leverage the following Pattern Language: intuitionmachine.gumroad.com/l/gpt4/6ndg9ax

1/n Maximizing profit margins and repeat customers. What's not to like about American capitalism? The idea is to get people hooked on a specific solution, and to milk that solution for what it's worth. What do you think the opioid crisis is all about? 🧵

https://twitter.com/EdbieLigerSmith/status/1890493554637566120

2/n You see it's not very far away from the consumer oriented attention economy that we've built. Companies are selling narratives. Religious institutions sell stories about salvation. Pharmaceuticals sell stories about paths to cures. Auto companies sell stories about physical mobility. The MIC sells stories about security. We've all bought in to narratives that we pay too much for at the expense of other narratives that companies cannot make a profit on.

1/n I am curious if you understand why o1-series models are prompted differently? 🧵 2/n A technical analogy would be that LLMs are prompted in an imperative manner and LRMs are prompted in a declarative way. It's similar to the distinction between working with Python and working with SQL (or a functional language). The latter does a lot of stuff underneath the covers that should not need to be expressed.

1/n Humans do not truly know if they are hallucinating. Even the best mathematicians have errors in their proofs. Ramanujan who derived infinite series equations through intuition was astonishing correct often, however some of his equations were in error after they were rigourously examined. 🧵 2/n You see, creativity is often stiffled by rigor. You don't want to be writing and editing at the same time. We are creative in the flow state and that flow state is predominantly driven by intuition. But to be effective in a flow state, we must have spent a lot of time in preparation.