AI-Augmented Software Development Manifesto

We are learning better ways of developing software with AI by doing it and helping others do it.

Through this work we have come to value:

Human agency and taste over agentic autopilot
Tight human feedback loops over long unattended agent runs
Reliable, tested software over abundant AI output
Clear, adaptable intent over rigid specs or vibe-coding

That is, while there is value in the items on the right, we value the items on the left more.

Principles

Ship real software, not AI theater. LOC count, agent swarms, and token burn are not progress unless users get something reliable.
Let requirements harden from evidence. Use AI to discuss, prototype, and test assumptions before locking the plan.
Treat code as cheap and understanding as expensive. Deliver increments quickly, throw away bad first passes, rewrite, and refactor aggressively.
Keep humans and AI in a tight loop. AI generates options quickly; humans supply taste, domain judgment, trade-offs, and the ability to say "no."
Build around people with taste and ownership. They should set clear boundaries, trust AI to execute within them, and have the standards to redo, undo, or throw away the result.
Prefer dialogue over prompt bloat and rigid process. Talk through the problem, agree on direction, then write down only what needs to persist.
Reliable, tested software is the measure of progress. Abundant code, docs, prompts, traces, and tool output are useful only when they improve the system.
Sustainable pace is non-negotiable. AI makes nonstop work seductive: late-night agent babysitting, context spirals, and fake urgency. A process that burns people out or degrades judgment is broken.
Technical excellence matters more when generation is cheap. Understand the architecture, probe the output, push back hard, and review security before plausible broken systems become production systems.
Simplicity is the art of refusing generated complexity. AI can generate code, dependencies, abstractions, and documentation faster than you can own them.
Good systems emerge from iteration and pruning. Prototype, learn, rewrite, document, and refactor. Do not expect unsupervised agents to develop taste or conviction.
Use AI to amplify agency, not replace it. It should help you learn faster, explore further, and spend more time where your taste and judgment have the highest return. If it makes you passive, you are using it wrong.

Author: Ronan Berder
Source: hunvreus/ai-manifesto
Original: Manifesto for Agile Software Development