AI-Augmented Software Development Manifesto

Takeaways

• Values human agency/taste, tight feedback loops, reliable tested software, and adaptable intent over autopilot, long unattended runs, output volume, and rigid specs.
• Ship real software—not LOC, swarms, or token burn; requirements should harden from evidence via prototype and test, not upfront lock-in.
• Code is cheap; understanding is expensive—rewrite, refactor, and throw away bad first passes aggressively.
• Sustainable pace matters: late-night agent babysitting and context spirals degrade judgment; simplicity means refusing generated complexity.
• Technical excellence rises in importance when generation is cheap—review architecture and security before plausible broken systems reach production.

Notes

Compact north-star for agentic vault and coding work—pairs well with eval-driven AGENTS.md discipline and against “vibe-coding” defaults in long agent runs.