Topic hubs · public-safe map
Series pages, not loose posts.
Notes are useful one at a time. Series pages make them usable as a map: what to read first, what is proven, what is still on the bench, and which downloadable kit belongs with the job.
Practical Agent Safety
How to put agents near real work without pretending a system prompt is a fuse.
- tool boundaries
- human approval gates
- verifiers
- rollback paths
Local-First AI Control
Use cloud models as useful workers while local policy owns context, logging and the stop path.
- kill switches
- witness models
- offline fallbacks
- private memory
Prompting as Specification Writing
Prompts that behave like job cards: scope, acceptance criteria, forbidden actions and evidence.
- work orders
- STOP gates
- review loops
- handoffs
The public map shows the working discipline, not the private wiring diagram. No internal infrastructure, customer details, credentials, or hidden operational routes. Enough to help a serious operator think better; not enough to expose John's bench.
Open the downloadable kits →