Claude Code Just Became Self-Aware

SUMMARY

Craig Hewitt demonstrates implementing self-reinforcing learning loops in Claude Code, inspired by the Compound Engineering plugin, to enhance AI agents for redesigning his personal website with Next.js and Sanity, making the system smarter over time.

STATEMENTS

The biggest advance in AI has been its ability to get smarter independently, particularly through agentic platforms like Manis and Claude Code over the last year.
Continuous learning feedback loops are essential for AI systems to improve without users repeatedly providing instructions.
The Compound Engineering plugin from every.to enhances Claude Code for development by incorporating compounding feedback and iterative learning, though it consumes many tokens.
Open-source nature allows customization of concepts from plugins like Compound Engineering to fit specific needs without full installation.
Redesigning a personal website using Claude Code with Next.js and Sanity avoids clunky tools like Framer by enabling conversational AI-driven changes.
Implementing an iterative learning loop involves skills triggered on code commits to update preferences and reduce future mistakes.
Plan mode in Claude Code facilitates thinking and iteration before execution, ensuring smarter planning with loaded context from past learnings.
Learnings should be stored project-specific in markdown files to persist preferences without cross-contaminating other projects.
Post-commit hooks can prompt users to capture worth-remembering insights, building searchable institutional knowledge.
Knowledge management is crucial to prevent unbounded growth in learning files, using compaction skills to maintain reasonable sizes for context loading.

IDEAS

AI agents like Claude Code can evolve into self-aware systems by embedding self-reinforcing feedback loops that learn user preferences automatically.
Open-source plugins provide blueprints for lightweight custom implementations, allowing users to extract only desired features like iterative learning without overhead.
Conversational interfaces with AI enable non-experts to direct complex development tasks, such as website redesigns, by simply describing desired outcomes.
Persistent memory in AI projects counters the limitation of session-bound context, creating compounding intelligence that improves with each interaction.
Plan mode acts as a strategic upfront thinking phase, loading historical learnings to infuse intentionality into AI actions and minimize errors.
Post-commit prompts transform routine coding into ongoing education for the AI, capturing nuances like style preferences in real-time.
Functional categorization of learnings, such as design or backend, prevents context overload and supports scalable knowledge bases.
Automatic compaction of knowledge files ensures long-term usability, merging duplicates and archiving old entries to fit LLM context windows.
Directing AI as an "agent director" involves high-level guidance rather than micromanagement, leveraging the system's superior problem-solving.
Cross-pollination from industry-standard tools like Compound Engineering accelerates personal AI setups, democratizing advanced capabilities.

INSIGHTS

True AI advancement lies not in raw intelligence but in self-improvement mechanisms that adapt to individual workflows, fostering exponential efficiency gains.
Customizing open-source AI enhancements empowers users to tailor tools precisely, avoiding bloat while amplifying core benefits like persistent learning.
Effective AI collaboration thrives on upfront context loading, where historical preferences guide planning to produce more aligned, error-free outputs.
Knowledge persistence in AI systems must balance accumulation with curation to avoid diminishing returns from context saturation.
User-AI symbiosis evolves when humans act as directors, providing vision while AI handles execution, optimizing for human strengths like intuition.
Iterative feedback loops in development tools mirror human learning, turning repetitive tasks into opportunities for systemic refinement and mastery.

QUOTES

"The biggest advance in AI has been for it to get smarter by itself."
"Each unit of engineering work should make subsequent units easier, not harder."
"I am a director of agents. That's all that I am."
"Memory and context management is the name of the game when it comes to any AI system and especially agents because they should be getting smarter and smarter over time."
"Context in Claude Code is limited to a session or if you clear context like it's gone right now we're persisting this memory forever in a way that's manageable."

HABITS

Using AI tools like Claude Code and Manis as daily drivers for nearly all tasks to streamline workflows.
Employing voice-to-text tools such as Super Whisper for natural interaction with AI systems during development.
Iterating in plan mode before execution to build intentionality and reduce mistakes in AI-assisted projects.
Prompting AI post-commit to capture learnings, ensuring continuous refinement of preferences and styles.
Reviewing and customizing open-source resources to fit personal needs, avoiding full installations of heavy plugins.

FACTS

AI agents and platforms like Manis and Claude Code have driven self-improving AI capabilities primarily in the last year.
The Compound Engineering plugin significantly boosts Claude Code but consumes substantial tokens, making it unsuitable for lighter use cases.
Context in tools like Claude Code is typically session-limited unless explicitly persisted through mechanisms like markdown files.
Recent data suggests optimal AI context files should remain under a certain length to avoid overwhelming LLM memory limits.
Open-source ecosystems enable rapid adaptation of advanced features, such as compounding learnings, into custom projects without proprietary dependencies.

REFERENCES

Compound Engineering plugin by every.to for iterative learning and feedback loops.
Claude Code platform for agentic AI development with hooks and skills.
Next.js framework integrated with Sanity CMS for website building.
100 Days of AI video series on daily AI explorations.

HOW TO APPLY

Analyze an existing open-source plugin like Compound Engineering by pasting its repository link into Claude Code's plan mode to extract core concepts such as continuous learning.
Define project-specific storage for learnings in a dedicated folder, using markdown files categorized by patterns, solutions, and preferences to maintain organization.
Set up a post-commit hook in Claude Code that prompts after each code commit, asking if anything is worth remembering to capture user insights immediately.
Implement a recall skill that automatically loads relevant learnings into context during plan mode, ensuring AI incorporates past preferences before starting tasks.
Establish knowledge management by limiting files to 50 entries, invoking a compaction skill to merge duplicates and archive outdated information for sustainable growth.

ONE-SENTENCE TAKEAWAY

Implement self-reinforcing learning loops in Claude Code to create persistently smarter AI agents tailored to your development preferences.

RECOMMENDATIONS

Customize AI tools by extracting lightweight features from heavy plugins to suit your workflow without token waste.
Always use plan mode for complex tasks to load historical context upfront, enhancing AI accuracy and alignment.
Adopt post-commit prompts habitually to build a growing knowledge base that reduces repetitive explanations over time.
Categorize learnings by functional areas from the outset to prevent context bloat and improve retrieval efficiency.

MEMO

In the rapidly evolving world of artificial intelligence, Craig Hewitt, a tech enthusiast and content creator, is pushing the boundaries of AI-assisted development. Frustrated with the limitations of drag-and-drop website builders like Framer, Hewitt turns to Claude Code, an agentic AI platform, to redesign his personal site using Next.js and Sanity as a content management system. What sets his approach apart is not just the tools, but the ingenuity of embedding self-reinforcing learning loops—mechanisms that allow the AI to remember and adapt to his preferences with every interaction.

Hewitt draws inspiration from the open-source Compound Engineering plugin, developed by the team at every.to, which supercharges Claude Code with iterative feedback for robust development. Though he finds the full plugin token-heavy for his needs, he masterfully extracts its essence: the concept of compounding knowledge. By conversing naturally with Claude Code—using voice tools like Super Whisper—he instructs the AI to analyze the plugin's repository and devise a streamlined version. This involves creating skills and hooks that activate on code commits, prompting Hewitt to note valuable insights, such as stylistic choices or problem-solving patterns, which are then stored in project-specific markdown files.

The beauty of this system lies in its persistence. Unlike standard AI sessions where context evaporates, Hewitt's setup builds a searchable institutional memory. In plan mode, the AI recalls past learnings before diving into tasks, ensuring fewer mistakes and more aligned outputs. He emphasizes managing this knowledge wisely—capping files at 50 entries and using compaction skills to merge redundancies—preventing the overload that plagues large language models. As Hewitt quips, he's merely a "director of agents," guiding the AI's superior execution with high-level directives.

This innovation isn't isolated; it reflects a broader shift toward AI that learns like a human collaborator, growing smarter through use. For Hewitt, who relies on platforms like Manis and Claude Code daily, the result is an "unfair advantage" in 2026's tech landscape. By sharing his process openly, he invites others to replicate it, democratizing advanced AI capabilities. As he continues this series—building on his "100 Days of AI" marathon—viewers witness not just a website redesign, but a blueprint for future-proofing personal and professional workflows.

Ultimately, Hewitt's experiment underscores a profound truth: AI's true power emerges when it evolves beyond one-off tasks into a compounding partner. In an era where memory management defines effective agents, his method offers a practical path to harness that potential, turning frustration with clunky tools into seamless creation.