ИИ-разработка для настоящих инженеров

This information could be useful for someone (Lesson 48):
Refined Questions and Answers
Q: How do you ensure an agent follows Domain Driven Development (DDD) and best practices if you are against large skills? A: Encode your software practices into every skill you use. If you care about a specific architectural concept, ensure it is mentioned in planning phases. Additionally, consider using a cheaper model for implementation followed by an "expensive" agent for a dedicated review stage against your specific standards.
Q: Would you recommend using Test-Driven Development (TDD) to create only the code needed for expected behavior? A: Yes. TDD and the "red-green-refactor" loop are excellent for AI because agents thrive on the immediate feedback loops that tests provide.
Q: Do you have a good setup for sharing coding standards via skills between different repositories? A: Not currently, as the speaker primarily works solo on a few projects. However, he suggests that standards could be shared via an npm package with a post-install hook to integrate them into projects.
Q: How do you keep your environment tidy and avoid mixing up the wrong skills as your collection grows? A: You must be "paranoid" about your context window. Be vigilant about removing unused skills and wordy descriptions that don't need to be there, as they compete for the agent's attention.
Q: Do you worry about losing your ability to code, and how do you prevent that? A: No. The speaker distinguishes between tactical coding (short-term firefighting) and strategic programming (long-term design and architecture). AI has "eaten" tactical programming, but it is not yet good at strategic thinking. He focuses on the strategic side, which he believes is the most important skill for developers in the AI era.
Q: If expanding the context window of newer models just adds to the "dumb zone," what is the point? A: Massive context windows (like 1 million tokens) are partly for marketing. While the "smart zone" might eventually creep up, a large stack allows companies to charge more for tokens. The speaker remains skeptical of the practical utility of extremely deep context windows.
Q: Do you evaluate new models (like a new Opus version) before using them in your workflow? A: No formal evaluation is performed. The speaker treats model evaluation like employee evaluation; he builds processes that allow any capable model to succeed and prefers to switch between providers quickly rather than being tied to one.
Q: What is your approach to writing skills and knowing when to create a new one? A: Create a skill when you find a gap in your process that requires a reusable command. Skills should be kept as concise as possible because they are easier to evaluate, security check, and change.
Q: How should one best manage multiple parallel agents? A: Tools like Sound Castle are designed for this, allowing you to use work trees to let multiple agents "plow through work" simultaneously.
Q: How do you handle context across a workflow involving multiple humans and AI tools? A: Generate handoff artifacts at every stage. For example, summarize a meeting transcript into a PRD. Use context pointers (via MCP servers or GitHub issues) so the agent can fetch "primary sources," such as original conversation transcripts, only when needed.
Q: Should you perform "Grill," "PRD," and "Issues" in one session or separate them? A: If your context window allows, do them in one session. This keeps the "primary source" knowledge from the grilling session alive for the creation of the PRD and subsequent tickets.
Q: What are the best strategies for tackling the "dumb zone" in medium to large projects? A: It is not about project size but whether a codebase is easy or hard to explore. You can make a large codebase AI-friendly by using a shared language (so the AI knows what terms to search for) and "deep modules" (fewer, larger chunks of logic).
Q: What is the difference between using a "handoff" skill and the "branch/fork" feature in an agent? A: Handoff is more flexible. It creates a markdown file that can be passed to different agents (e.g., from Claude to Codex) or even saved as a GitHub issue. Branching is restricted to the same project and agent.
Q: Should I use a cheaper agent for AFK tasks to optimize for cost? A: No. You should optimize for quality. If you reduce quality to save money, you will likely spend more of your own expensive time fixing agentic mistakes than you saved on token costs.
Q: What is the roadmap for absolute beginners with no coding background? A: Beginners should find a "mission" they care about. The speaker recommends using his "Teach" skill, which encodes 10 years of teaching experience to guide learners through the "zone of proximal development".
Q: Should I use AI to write my skills, or write them manually? A: Use AI for a first pass to generate the basic idea, but then manually check every line to remove duplication and ensure conciseness.
Q: How can we meet career progression expectations if managers see our work as "just AI"? A: Focus on and talk about your strategic skills. Explain that the AI handles the tactical "heavy lifting" while you are responsible for the strategic direction and building the "outer harnesses" that make the AI productive.
Q: Does the "memory.md" file managed by Claude Code help or clutter the context? A: The speaker hates it and recommends deleting it. He prefers agents to be stateless in a codebase for consistency and wants full manual control over everything entering the context window.
Q: How is AI changing the software development life cycle (SDLC)? A: Implementation time is trending toward zero. This means the SDLC will involve much more time spent on planning and QA, as those become the primary bottlenecks now that execution is automated.
Q: How do you use AI for your own learning and research? A: Interestingly, the speaker performs research without AI by reading books and taking notes in an Obsidian graph. He prefers to keep his learning intentional and personal to avoid being flooded with AI-generated "guff".
Q: What is your take on the "hammer and nail" scenario where everything becomes a skill or AI-first? A: AI is a "Golden Hammer." It is worth trying to apply it everywhere because it can turn "immersive" human tasks into "process" tasks (AFK) that run in the background, saving significant time.
Q: What are the best strategies to make a pre-AI codebase "AI ready"? A: Good design for humans is good design for AI. Focus on making the codebase "easy to change" by adding test seams, improving feedback loops, and ensuring automated checks are in place.

This information could be useful for someone (Lesson 47):
Refined Questions and Answers
Q: What has been the biggest change in the world of AI that has changed my workflows over the past month or two? A: The speaker has transitioned to using GitHub as a first-class runner for agents. Instead of purely local development, he uses Sand Castle (https://github.com/mattpocock/sandcastle) inside GitHub Actions to trigger agents on specific issues and tasks.

Q: What is the benefit of running agents in the cloud versus running a harness locally? A: Cloud runners (like GitHub Actions or Kubernetes) are superior for AFK tasks because they provide isolated sandboxes. This eliminates the need for the user to manually manage multiple work trees on their local machine.

Q: How do you determine when to use "Human in the Loop" (HITL) versus triggering AFK workflows? A: HITL is best for planning, interviewing (grilling), and reviewing. AFK is used for execution—delegating well-specified, repetitive, or longer tasks to scale personal productivity by running multiple background jobs simultaneously.

Q: Can you describe your thought process when writing "skills," including best practices? A: The most effective skills are short and concise. Long skills create "noise" that competes for the agent's attention. The speaker recommends deleting duplication and useless words, treating skill writing like using a "chisel or a scalpel" for precision.

Q: Can you explain some of the first principles of good software design? A: A good codebase is defined as one that is "easy to change." This is achieved by designing systems where changes in one area do not ripple out to others and by maintaining automated tests to verify stability.

Q: Do we in cohort 4 get access to future cohort material? A: Generally, you are purchasing access to the specific cohort you signed up for. While some previous students were granted access to this session, future updates are not guaranteed.

Q: What are the most effective methods for describing how code should be written? A: Avoid micromanaging the agent's style unless it causes a "material issue" like a bug or makes the code harder to change. It is better to widen the "bounds of acceptability" for the AI.

Q: Where is the best place to add steering instructions (planning, implementation, or review)? A: The review stage is the best location. Adding instructions during planning is too high-level, and putting them in global files like claw.md bloats every session unnecessarily.

Q: What has been your biggest "aha" moment while putting together the latest course material? A: The realization that "plan mode" is often detrimental. The speaker replaced earlier lessons praising plan mode with content explaining why it "sucks" compared to more modern approaches.

Q: Have you tried any other agentic harnesses, and what makes you keep coming back to Claude Code? A: The speaker uses Claude Code primarily because it is currently integrated into his workflow, though he builds tools like Sand Castle to be harness-agnostic.

Q: What is your view on the recent change Anthropic made regarding OAuth tokens? A: The speaker views the move from setup tokens to OAuth as a significant cut to the functionality of AFK agents, despite how it was marketed by the company.

Q: What do you do if you find your context window has too many tokens (the "dumb zone")? A: You have two primary options: Handoff (transferring the task to a fresh session for prototyping) or Compact (summarizing the current session history to reduce token count and improve the model's focus).

Q: What is the most high-leverage AI coding skill in your opinion? A: The "Grill Me" (or "Grill with Docs") skill is considered the most impactful for interviewing the AI and refining technical ideas.

Q: Should I have docs of patterns in a specific project folder or in global rules? A: Use "progressive disclosure." Instead of dumping everything into global rules (which bloats the context window), use context pointers to link to local markdown files only when needed.

Q: What do you think about AI reviewing production PRs without humans in the loop? A: Automated AI review is an essential tool. It should be used to catch obvious errors and security issues so the human reviewer can focus exclusively on final, high-level checks.

Q: Is Sound Castle making any adjustments for the upcoming Claude licensing changes? A: Yes, adjustments will be made once the specific mechanisms of the licensing changes are finalized.

Q: Why is the "Grill Me" skill so impactful despite being brief? A: Short skills are easier to maintain and produce more focused output. The speaker compares it to the Gettysburg Address—impactful because of its brevity and precision.

Q: Is there a real benefit to using memory.md files versus claw.md? A: The speaker distrusts automated memory systems. He argues that agents should be stateless in a codebase; stateful agents become bloated and harder to optimize over time.

Q: What is the difference between using a "Goal" and a "Ralph loop" in Claude Code? A: The "Goal" feature is wasteful and expensive, often reaching 500,000 tokens. Ralph loops or AFK agents provide more control by slicing tasks into segments, keeping the model in the "smart zone" and reducing costs.

Q: Have you tried coding with alternative models like Kimmy or Mistral? A: No. The speaker believes proprietary models like Claude Opus are still significantly more capable for coding tasks and is not currently focused on cost-saving via open-weight models.

Q: What are your top MCP servers and how do you fetch documentation? A: The speaker does not use MCP servers for documentation. He recommends caching research into local markdown files so the agent can reference them as specific assets only when needed.

Q: When you're learning something genuinely hard and new yourself, what's the method you use? A: He uses a custom "Teach" skill based on 10 years of teaching experience. It balances knowledge, interactive skills, and community wisdom while ensuring the user stays in the "zone of proximal development".

Q: Do you think code quality matters anymore? A: It matters more than ever. AI agents behave like "new contractors" every time they enter a codebase; if the code is coupled or messy, the agent will struggle to understand and improve it.

Q: Can you explain and demonstrate Git basics like reset, cherry-pick, and pull? A: These specific technical demonstrations are covered in a separate dedicated video within the course content.

:smile

The course materials only html files in the code.zip , wasnt supose to have the code files with the skills?

I think these are the links you're looking for:

https://github.com/mattpocock/skills
https://www.aihero.dev/skills

If I remember correctly, Matt has his own 'skill' project; you can simply search for it on GitHub.

Alternatively, if he used them during the course, you ought to find them within the lectures' HTML files.

Как курс? Начал смотреть что-то не понимаю пока ценности.

Matt Pocock очень очень крутой специалист в этой теме. Я сам еще не смотрел этот курс, только начинаю, но выбрал сам курс именно из-за этого человека

Если не до конца ясно зачем все это нужно, советую просто вбить в ютубе его имя и найти лекции\интервью около часа ( в 10 раз меньше чем тут в курсе) чтобы понять идею того о чем он говорит. А идея такова что он пытается научить людей эффективнее работать в паре с ИИ. Понимать что делать, видеть системный дизайн продукта, понимать как поступать в тех или иных ситуациях как разработчик.

Do I understand correctly that this course is essentially an updated version of the previous Cohort 003 course, “Claude Code for Real Engineers”?

Or are these two completely different courses with different content, goals, and curriculum?

After comparing both courses, it looks like this is essentially an updated or "second edition" of Claude Code for Real Engineers that incorporates the latest improvements Matt has made to his workflow. So, I would skip Claude Code for Real Engineers and take this course instead.

same course. this bald guy wants to make more money by teaching the same shi using the cohort format in different titles.

This is such a bad take. He's publicly explained that he changed the title because the scope of the course changed. It focuses on fundamental concepts and a development workflow that's agent agnostic, so calling it a "Claude Code course" no longer made sense. Also, the old version isn't even being sold anymore, and everyone who bought it got the new version for free. That makes your argument even less convincing.

There's a file in the course materials called Appendix: Cohort 003 → Cohort 004 Changelog where you can see all the changes compared to the previous cohort.

Update: Final Office Hours

Archive Added

Course finished 👏

Thanks again! I really appreciate it.

Is this complete?

Can't find the Download course archive button. why?

course is still in progress after it is completed only then download is available, but this is not 2016, in few weeks information will be obsolete, so watch it here, apply it, and move on, download just gives you dopamine hit like you got something from it, you got nothing by downloading it. thanks for advice geronimo. no problem.

Update: Day 4 - Day 6 + Office hours

Two office hours left, this Friday!

thanks!

Hello team! Can you please add the latest lessons! Thanks in advance!

Thanks for making this cohort a separate post. For a moment, I got worried that the previous cohort had been replaced and I had lost all my watch progress. :/

Why doesn't the course has download course archive ?

it is still on going and it will end June 12th. Once it ends there will be download possibility.

OK, Thanks for your answer

You guys rock! Thank you!

Администрация, Ваш впн навернулся, если что, это правда. С позавчера не работает.

Попробуйте сейчас

вот проверил несколько раз, на разных устройствах, на разных клиентах, нет, пока что нигде не работает. к сожалению

Попробуем разобраться сегодня. Возможно vpn попал под блокировку в РФ.

да, я просто хотел вас предупредить. спасибо заранее, если что

Спасибо 🤝

“Claude Code for Real Engineers” now outdated and this is a new version? Or better to pass both?

There's a file in the course materials called Appendix: Cohort 003 → Cohort 004 Changelog where you can see all the changes compared to the previous cohort.

👉 Курс в процессе. Новые уроки добавляются по мере готовности. Окончание потока — 12 июня. Архив будет доступен после завершения курса.

👉This course is currently in progress. New lessons will be added progressively, and the cohort concludes on June 12th.
The course archive will be made available after the final release.

Wow, you guys are amazing. Thanks so much!

Dude I really appreciate your hard word! Thanks so much!