Articles
-
GuidesContext Engineering for Internal Assistants
Internal assistants fail less when context is assembled deliberately instead of dumped wholesale into the model. Here is a practical context-engineering approach: task framing, scope, retrieval, compression, and provenance.
-
GuidesPrompt Caching for LLM Apps: Where It Actually Pays Off
Prompt caching only pays when your reusable prefix is stable, versioned, and safe to share. The hard part is not turning it on; it is deciding what may be cached.
-
GuidesMCP Servers in Production: Start Narrow, Stay Auditable
MCP works when you treat servers as trust boundaries, not generic adapters. Narrow resource scope, explicit consent, and auditability matter more than broad connectivity.
-
OpinionWhat Belongs in the System Prompt vs the App Layer
System prompts are for steering behavior. Authorization, state changes, retries, and data access belong in the app and server layers where they can be enforced.
-
GuidesWhat to Log for LLM Apps Before You Need It
A concrete logging model for LLM apps: traces, tool calls, approvals, versioned context, and the minimum metadata needed to reconstruct failure.
-
GuidesHuman Approvals for Agents: Where to Put the Breakpoints
A concrete guide to approval boundaries for agents: side effects, trust boundaries, ambiguous decisions, and the audit trail that makes the approval real.
-
GuidesQueue Design for Long-Running Agents
How to design durable queues for agent jobs: state machines, idempotency, checkpoints, cancellation, and the failure modes SQS and Temporal force you to handle.
-
OpinionWhy Internal AI Assistants Fail at Authentication Boundaries
Internal assistants usually fail where identity, permissions, and retrieval meet. Here is why broad access feels productive at first, and why scoped identity design matters more than another prompt tweak.
-
GuidesWhen to Fine-Tune vs Retrieve vs Prompt
Fine-tuning, retrieval, and prompt engineering solve different failure modes. Here is the decision framework I use when a team asks how to make an LLM app more accurate, cheaper, or easier to operate.
-
GuidesHow Does OpenClaw Work? A Guided Tour of the Lobster Assistant
OpenClaw (formerly Clawdbot, then Moltbot) is a personal AI assistant that runs on your devices. Here is the architecture: the Gateway control plane, channels, sessions, tools, skills, and the agent loop.