Post

More from @IntuitMachine

Carlos E. Perez

@IntuitMachine

Oct 1

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

Read more about AI Agentic Patterns: intuitionmachine.gumroad.com/l/agentic/zo5h…

Read 4 tweets

Carlos E. Perez

@IntuitMachine

Sep 15

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.

New prompting patterns not explicitly documented in A Pattern Language for Agentic AI

🆕 1. Diff-and-Contextual Citation Pattern

Description:
Instructs agents to generate precise citations with diff-aware and context-sensitive formatting:

【F:†L(-L)?】
Includes file paths, terminal chunks, and avoids citing previous diffs.

Why It’s New:
While Semantic Anchoring (Chapter 2) and Reflective Summary exist, this level of line-precision citation formatting is not discussed.
Function:

Enhances traceability.
Anchors reasoning to verifiable, reproducible artifacts.

🆕 2. Emoji-Based Result Signaling Pattern

Description:
Use of emojis like ✅, ⚠️, ❌ to annotate test/check outcomes in structured final outputs.
Why It’s New:
No chapter in the book documents this practice, though it overlaps conceptually with Style-Aware Refactor Pass (Chapter 3) and Answer-Only Output Constraint (Chapter 2).

Function:

Encodes evaluation status in a compact, readable glyph.

Improves scannability and user confidence.

🆕 3. Pre-Action Completion Enforcement Pattern

Description:
Explicit prohibition on calling make_pr before committing, and vice versa:
"You MUST NOT end in this state..."

Why It’s New:
This kind of finite-state-machine constraint or commit-to-pr coupling rule is not in any documented pattern.

Function:
Enforces action ordering.

Prevents invalid or incomplete agent states.

🆕 4. Screenshot Failure Contingency Pattern

Description:
If screenshot capture fails:
“DO NOT attempt to install a browser... Instead, it’s OK to report failure…”

Why It’s New:
Not part of any documented patterns like Error Ritual, Failure-Aware Continuation, or Deliberation–Action Split.

Function:
Embeds fallback reasoning.

Avoids cascading errors or brittle retries.

🆕 5. PR Message Accretion Pattern
Description:
PR messages should accumulate semantic intent across follow-ups but not include trivial edits:

“Re-use the original PR message… add only meaningful changes…”

Why It’s New:
Not directly covered by Contextual Redirection or Intent Threading, though related.

Function:
Maintains narrative continuity.

Avoids spurious or bloated commit messages.

🆕 6. Interactive Tool Boundary Respect Pattern
Description:
Agent should never ask permission in non-interactive environments:

“Never ask for permission to run a command—just do it.”

Why It’s New:
This is an environmental interaction boundary not captured in patterns like Human Intervention Logic.
Function:

Avoids non-terminating agent behaviors.

Ensures workflow compliance in CI/CD or batch systems.

🆕 7. Screenshot-Contextual Artifact Embedding
Description:
Use Markdown syntax to embed screenshot images if successful:

![screenshot description]()

Why It’s New:
While there’s mention of Visual Reasoning in earlier books, this explicit artifact citation for visual evidence is not patterned.
Function:

Augments textual explanation with visual verification.

Supports interface-testing workflows.
🧩 Summary Table

Read 4 tweets

Carlos E. Perez

@IntuitMachine

Aug 9

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!

Just like in previous threads like this, I will use my GPTs (now GPT-5 powered) to break down the prompts in comprehensive detail.

Read 7 tweets

Carlos E. Perez

@IntuitMachine

Aug 5

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.

3/n Let's begin with illegal immigration and then I'll work the argument up to religion, the military, and finally the state.

Read 12 tweets

Carlos E. Perez

@IntuitMachine

Jul 5

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

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

More analysis from a dark triad perspective:

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

FYI. A quick primer on Spiral Dynamics:
medium.com/p/0ef0ceb1ff80

Read 8 tweets

Carlos E. Perez

@IntuitMachine

Jul 4

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.

3/n One universal commonality is that these are all "intuition machines," and they share the epistemic algorithm that learning is achieved through experiencing. Thus, all these systems (humans included) share a flaw of cognitive biases.

Read 12 tweets

Share this page!

Enter URL or ID to Unroll

Carlos E. Perez

Try unrolling a thread yourself!

More from @IntuitMachine

Carlos E. Perez

Carlos E. Perez

Carlos E. Perez

Carlos E. Perez

Carlos E. Perez

Carlos E. Perez

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!