Common Concepts & Insights
The ideas that cut across every system here — the shared pattern, the shapes of verification, where the safety boundaries actually live, and what building one of these costs.
The pattern & the promise
Orchestrator, subagent, tool; and the rule everything serves: they prepare, you decide.
✅Four shapes of verification
A bounded loop, a single adversarial pass, a computed tie-out, or no check at all — chosen per system by what a wrong answer costs.
🛡️Why the promise holds
The guarantees are tool grants, not good intentions.
📖A skill is not an agent
Command vs subagent vs skill, and why tools: lives on agents.
Context: two axes, not one ranking
Scope and timing decide what loads; precedence is a separate question for when what's loaded disagrees.
📚One offline knowledge base, six agents
A local wiki isn't a web tool — so read access to it doesn't reopen the no-web-tools boundary three of these agents were built to hold.
🔨Writing a new one, and what it costs
The conventions, and the price of hardening vs building.
🧱Running GSD on a Web Application
Fourteen findings on coding, code management, and deployment from one real build.
🏷️When the name is the only handle
Why the nai- rename would have touched forty files, and what settling names early would have prevented.
🛤️The golden path for web apps
Twelve decisions to walk before the first line of code, and the aspects waiting beyond them; reuse applied at generation time, not through a shared package.