Every day, artificial intelligence is opening new ways to build software. Its potential is enormous, but it does not remove the need to understand the code we write or to take responsibility for it.
With this manifesto, we affirm that a developer’s value is not measured by the amount of code they produce, but by their ability to understand it, question it, and stand behind it. AI is a powerful lever, but it must never become a substitute for human judgment.
We are discovering a new way to build software with language models.
Through this practice, we have learned to value:
In other words, we recognize the value of what is on the right, but we prioritize what is on the left.
AI is an accelerator, not a replacement for clear thinking.
We consider developers accountable for every line of code they submit, whether written by them or generated by an LLM.
Before any submission, code should be:
Code review is not there to discover what the author did not take the time to verify.
AI gives us speed. That speed should be reinvested in understanding, quality, testing, and design, never in multiplying code we do not fully control.
Do not outsource your judgment to a model. That judgment is exactly what makes you a developer.
Submit only code you can defend
A developer should be able to explain the important choices in the code they propose: its logic, trade-offs, limits, and rejected alternatives.
The goal is not to know every detail by heart, but to be able to justify what enters the codebase.
Treat every generation as a proposal
Code produced by an LLM is never a validated answer. It is a working hypothesis, just like an example found online or a suggestion from a teammate.
It must therefore be evaluated, corrected, or rejected before adoption.
Keep control proportional to impact
The more critical, complex, hard to reverse, or expensive to maintain a change is, the higher the expected level of understanding and verification should be.
Not all lines carry the same risk, but no line becomes reliable just because it was generated confidently.
Prefer mastered code over impressive code
A simple, well-understood, and appropriate solution is better than a sophisticated one that nobody truly masters.
AI should not encourage technical one-upmanship, premature abstraction, or unnecessary concepts.
Reduce before you add
The ease of generation should not lead to producing more code than needed.
Every addition creates future cost: reading, testing, maintenance, documentation, security, and evolution. Generating quickly does not make this cost disappear.
Write for the next reader
Code should not only work. Its intent should be accessible to the person who will review, modify, or diagnose it later.
Readability is not an aesthetic preference: it is a condition for collective accountability.
Take ownership of the validation work before review
A submission (pull request) should present work that is already understood, reviewed, tested, cleaned up, and clarified.
Review is there to challenge choices, detect blind spots, and improve a proposal, not to perform the author’s first round of verification.
Take on the same accountability, whatever the code origin
Accountability for a line belongs to the person who submits it, not to the tool that generated it.
“The AI wrote it” is neither an explanation, nor a justification, nor a mitigation of accountability.
Use the time saved to raise quality
The benefit of AI should not be measured only in output volume or execution speed.
Time saved should help us better understand the problem, examine edge cases, improve tests, simplify design, and document important decisions.
Be willing to discard a generation
A suggestion that is hard to understand, verify, or simplify should be abandoned, even if it appears to work.
The cost already invested in a conversation with an LLM never justifies integrating poorly mastered code.
Suggestions and discussions are welcome.
The goal is not to be anti-AI, but to promote responsible AI-assisted software engineering.
This work is licensed under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0).
https://creativecommons.org/licenses/by/4.0/