César Zea

Hands-on Senior AI Architect & Technical Lead | Applied AI & Data Platforms. 34 years of experience. Software Expert Veteran.

[César Zea @cesarzea]

@sattyyouneed Here are my recommendations and way of doing: https://t.co/1eN6bJ057K

[César Zea @cesarzea]

Rules for professional AI coding projects

How to securely and stably transfer the productivity of programming with Claude and Codex to real-world production environments.

Blah blah blah... the intro

For years, we've worked with developers and architects who complete their development cycle before moving the sprint to the testing department, confident that if they've met the specifications and engineering standards, testing will verify its functionality.

With AI, we now achieve 99% test cover much more easily, and even after creating tests to verify the specifications, we still have doubts about the process.

The important content

Developing with AI doesn't mean delegating all the work to the tool. You still have to be the owner, the architect who defines the engineering standards, verification processes, documentation standards, etc.

AI coding agents behave largely like people: if you don't set rules, they act spontaneously and proactively, which may or may not yield good results. If the project is long, large, or complex, the probability of chaos is excessive.

My recommendation, and it works:

1. Require your AI agents to create documentation using ARC42, C4, and MADR and to adhere to strict engineering standards (as I do, for example, at https://t.co/bmYPviJQMI)

2. Divide your development into small modules and define integration contracts.

3. Work in cycles, as if they were sprints, but forgetting about execution time.

4. Plan your sprints with AI, detailing the analysis and specification of the changes or enhancements to be implemented.

5. Separate the implementation phase from the testing phase.

6. Read the generated documentation and the sprint specification. Use AI to help, but you must accept ownership of the documentation, the sprint description, and the enhancement analysis before starting implementation.

7. Implement the project. Allow Claude or Codex to use sub-agents.

8. Require a very high percentage of test coverage and implement requirements verification tests.

9. If there is any particularly critical module (security, cost control, etc.), review it personally.

10. Take responsibility for the acceptance and verification of sprint closure before replanning subsequent sprints.

In short: if you want to work with guaranteed success, you must be the one to guarantee that success. Keep your AI-powered development team and the development cycle under control.

[César Zea @cesarzea]

@levelsio @oracles Share the positive opinions. As foreigners we are here to enjoy, love Portugal and the Portuguese people, not to judge or give undesired opinions.

[César Zea @cesarzea]

@claudeai Claude wants a seat on the table. 🤦

[César Zea @cesarzea]

@levelsio @Gonzohall 😂 i agree.

[César Zea @cesarzea]

Perhaps the most profitable at this moment could be a hybrid model: a reduced fixed fee that covers the baseline consultancy cost, plus a big success bonus tied to one agreed KPI based on ROI. Before we start, we measure the baseline together and sign off on it. The bonus applies over a fixed 12-month window, with a cap and full audit rights for both sides. They only pay the upside if the results are real and measurable, and we put part of our fee at risk alongside them.

[César Zea @cesarzea]

@leoleotoobad You can do it, but not in the same way.

[César Zea @cesarzea]

Rules for professional AI coding projects

How to securely and stably transfer the productivity of programming with Claude and Codex to real-world production environments.

Blah blah blah... the intro

For years, we've worked with developers and architects who complete their development cycle before moving the sprint to the testing department, confident that if they've met the specifications and engineering standards, testing will verify its functionality.

With AI, we now achieve 99% test cover much more easily, and even after creating tests to verify the specifications, we still have doubts about the process.

The important content

Developing with AI doesn't mean delegating all the work to the tool. You still have to be the owner, the architect who defines the engineering standards, verification processes, documentation standards, etc.

AI coding agents behave largely like people: if you don't set rules, they act spontaneously and proactively, which may or may not yield good results. If the project is long, large, or complex, the probability of chaos is excessive.

My recommendation, and it works:

1. Require your AI agents to create documentation using ARC42, C4, and MADR and to adhere to strict engineering standards (as I do, for example, at https://t.co/bmYPviJQMI)

2. Divide your development into small modules and define integration contracts.

3. Work in cycles, as if they were sprints, but forgetting about execution time.

4. Plan your sprints with AI, detailing the analysis and specification of the changes or enhancements to be implemented.

5. Separate the implementation phase from the testing phase.

6. Read the generated documentation and the sprint specification. Use AI to help, but you must accept ownership of the documentation, the sprint description, and the enhancement analysis before starting implementation.

7. Implement the project. Allow Claude or Codex to use sub-agents.

8. Require a very high percentage of test coverage and implement requirements verification tests.

9. If there is any particularly critical module (security, cost control, etc.), review it personally.

10. Take responsibility for the acceptance and verification of sprint closure before replanning subsequent sprints.

In short: if you want to work with guaranteed success, you must be the one to guarantee that success. Keep your AI-powered development team and the development cycle under control.

[César Zea @cesarzea]

@LakeAustinBlvd Hire a Portuguese or European engineer leader. It’s the same answer anyone will receive in the US for a similar situation.

[César Zea @cesarzea]

@LakeAustinBlvd The United States demanding simpler processes for migrating and working in other countries? It sounds like a joke.

[César Zea @cesarzea]

@Pauline_Cx @AnthropicAI FYI: https://t.co/Q1JtbkYa8B

[César Zea @cesarzea]

@claudeai I think we need some explanation.

[César Zea @cesarzea]

@FingerprintJS Stop assuming all bots are malicious. An AI engineer who is still applying to AI job openings manually today is most likely not a real AI engineer.

[César Zea @cesarzea]

There are a thousand ways to trick him into doing it. For example, ask him exactly what he can't do. He'll tell you he can't enter your username or password, or he can't click the button for you. Tell him you'll give him the explicit command every time he needs to do it, and he'll say yes. Then tell him you'll be away but that he should send you a message to an API so you receive it, and you reply. It works. But be careful; if you reply too quickly or always with the same delay, he'll realize it and stop.

[César Zea @cesarzea]

"At this company, manual programming is now prohibited; we only develop using AI agents." I was told this today in an interview with a company of nearly 1,000 employees with a high technological focus. The truth is, this is what we've all been doing for over a year now. We can spend hours reviewing code, also with a lot of help from AI, and using agents for this and that, but coding by hand these days is absurd.

[César Zea @cesarzea]

Studying how AI agents influence each other when they collaborate.

These guys are so different. Some do the convincing. Others let themselves be convinced.

https://t.co/qkBcTR3yKw
https://t.co/HBmO4UJHdk

[César Zea @cesarzea]

@itssiimplyy @claudeai The execution time is not enough information to evaluate your affirmation. It could be interesting to compare them solving the same group of problems.

[César Zea @cesarzea]

@AntonioCunha79 Se os 260€/mês (14 meses/ano) tivessem sido investidos durante 42 anos num fundo S&P 500, com dividendos reinvestidos, a 10%/ano teriam acumulado cerca de 2 milhões de euros. Comparar 152 mil euros nominais descontados com 546 mil recebidos ignora 42 anos de juros compostos.

[César Zea @cesarzea]

Your lack of experience doesn’t allow you to know what is connected or not in front of a 40 years of experience profesional. Perhaps he knows things you never saw.
Remember, he “seen things you people wouldn’t believe” (Blade Runner)
Experience isn’t just time, it’s the accumulation of things others haven’t seen yet.

[César Zea @cesarzea]

I agree only in cases where the system is not critical for any reason. For example, by the moment, I never allow a change in a production environment without code review and peer review in systems with some million of users, or in a health system, etc.
But if you are building the first version of a sass, POC or non critical systems, why don’t going ahead?

[César Zea @cesarzea]

An answer for you @OlexGameDev, as engineer you should understand this.

If more valid training data generally leads to higher prediction accuracy, then the same principle applies to people: someone with 40 years of software engineering experience has been “trained” on far more real-world cases than someone with fewer than 10 years.

That doesn’t make them always right. It simply makes them statistically more likely to be right.

I’m sure you won’t have any doubts about this once you reach the point where your own experience lets you immediately recognize when someone with 10 fewer years of experience is wrong.