Four Values
We are constantly discovering new ways to build products with AI, doing it ourselves and helping others. Through this work, we have come to value:
Psychological safety and tacit knowledge of the team over adopting trendy AI tools and metrics.
Thorough planning and rigorous verification over speed of code generation and blind "vibe coding."
Asynchronous management of multiple AI agents over traditional focused synchronous development.
Voluntary bottom-up experiments and sandboxes for enthusiasts over top-down AI mandates enforced by threat of termination.
That is, while there is value in the items on the right, we value the items on the left more.
Eight Principles
1. Planning pays off tenfold
One minute spent on a detailed spec for an agent saves 10 minutes of debugging cascading bugs. An agent can work autonomously for 1.5 hours - but only with a quality technical specification.
2. AI is a catalyst, not a savior
If the team had overload, distrust, and blurred roles before AI adoption - AI will amplify this chaos at incredible speed. Fix organizational problems first, then implement technology.
3. Humans are the bottleneck
When accelerating code generation with agents, remember the Theory of Constraints: the next stage (review) receives exponentially more tasks. AI speeds up code generation 3x - reviewers get 3x more Pull Requests. Without end-to-end metrics (Lead Time: from idea to production), reviewers burn out.
4. Protect the holders of tacit knowledge
AI agents cannot extract from code or documentation what isn't there. Business context, reasons behind architectural decisions, political agreements - all of this lives in specific people's heads. Their task is to methodically convert this context into formalized specs for AI agents.
5. Accountability comes only from voluntary choice
Management pressure and KPIs for AI usage lead to hidden sabotage (31% of employees) and formal reports. Create sandboxes, find enthusiasts, give them tools and time. Rely on organic growth through AI champions.
6. Writing code is dying, long live engineering
The classic SDLC is transforming. Stages collapse into a single agent window. Engineering work migrates to the "edges": requirements preparation (input) and rigorous validation of results (output). Planning + Verification > Coding.
7. Teams learn only through their own practice
Creating perfect prompts by external consultants doesn't work. The tool becomes an artifact that nobody uses. Teams must build their own experience, maintain a shared knowledge base, and exchange prompts internally.
8. We accept the "Productivity Paradox"
AI adoption alone does not increase overall business productivity. If the manager doesn't restructure processes, time saved by agents won't convert into results. Process yields a derivative. Tools yield only a fixed increment.
What to do about it
- Legitimize sandboxes. Find enthusiasts, give them tools and time, but not new KPIs.
- Switch to end-to-end metrics. Measure Lead Time from idea to production, not code writing speed.
- Ensure transparency. Implement automatic export of AI work sessions for review.
- Watch the team's energy. Don't implement AI during overload. Engage through voluntary choice.