Notes on AI-assisted software development. Tooling, workflow, and team practices for building with AI agents.https://patrickelmore.com/en-ushttps://patrickelmore.com/posts/when-letting-it-rip-goes-wrong/https://patrickelmore.com/posts/when-letting-it-rip-goes-wrong/A year of disciplined Claude Code use, sophisticated guardrails, more than a thousand hours of practice. The agent destroyed my database anyway.Sun, 17 May 2026 00:00:00 GMTclaude-codefield-notesincident-reportai-developmenthttps://patrickelmore.com/posts/generation-is-not-retrieval/https://patrickelmore.com/posts/generation-is-not-retrieval/Developers treat hallucination as a lookup that went wrong. The mechanism is that the model generates plausible continuations and there is no lookup. Verification isn't a bug to fix on the model side. It's a structural job on the user side.Sat, 16 May 2026 00:00:00 GMTmental-modelsllmhallucinationprompt-engineeringai-developmenthttps://patrickelmore.com/posts/attention-is-not-memory/https://patrickelmore.com/posts/attention-is-not-memory/The model doesn't remember earlier turns. It re-reads the full transcript from scratch every turn. What feels like memory is re-reading, and the failure modes follow directly from that.Fri, 15 May 2026 00:00:00 GMTmental-modelsllmcontext-windowprompt-engineeringai-developmenthttps://patrickelmore.com/posts/context-window-is-the-workspace/https://patrickelmore.com/posts/context-window-is-the-workspace/The folk model treats the context window as storage you load before the model knows things. The mechanism is the opposite: the window is the workspace itself, and every token in it shapes every answer the model gives.Sat, 09 May 2026 00:00:00 GMTmental-modelsllmcontext-windowprompt-engineeringai-developmenthttps://patrickelmore.com/posts/working-with-llms-effectively/https://patrickelmore.com/posts/working-with-llms-effectively/Effective collaboration with a capable LLM is mutual deference based on honest capability assessment. Most prompt engineering is the human refusing to do the partition.Thu, 07 May 2026 00:00:00 GMTmental-modelsllmprompt-engineeringai-developmentagent-workflowhttps://patrickelmore.com/posts/feedback-loops-ai-agents/https://patrickelmore.com/posts/feedback-loops-ai-agents/Capture is cheap and aggressive. Promotion to rules is expensive and conservative. The discipline that separates compounding feedback from rule-set pollution is the gate between them.Sat, 02 May 2026 00:00:00 GMTengineering-with-agentsfeedback-loopsagent-workflowai-developmentsoftware-processhttps://patrickelmore.com/posts/multi-agent-orchestration-patterns/https://patrickelmore.com/posts/multi-agent-orchestration-patterns/Once more than one agent is in a pipeline, the load-bearing engineering act is the discipline at the boundaries between them.Wed, 29 Apr 2026 00:00:00 GMTengineering-with-agentsmulti-agent-orchestrationagent-workflowai-developmentsoftware-processhttps://patrickelmore.com/posts/composing-the-context-agent-failures/https://patrickelmore.com/posts/composing-the-context-agent-failures/Most failures attributed to bad model output are composition failures: the wrong material, in the wrong order, with no rule for which source wins.Tue, 28 Apr 2026 00:00:00 GMTengineering-with-agentscontext-engineeringagent-workflowai-developmentsoftware-processhttps://patrickelmore.com/posts/context-pollution-debris-not-capacity/https://patrickelmore.com/posts/context-pollution-debris-not-capacity/Output quality degrades because the room is full of debris, not because the room is full. The fix is structural, not a bigger window.Mon, 27 Apr 2026 00:00:00 GMTengineering-with-agentscontext-engineeringagent-workflowai-developmentsoftware-processhttps://patrickelmore.com/posts/tiered-planning-engineering-discipline/https://patrickelmore.com/posts/tiered-planning-engineering-discipline/Iterative development applied to the planning function itself. The checkpoints between tiers are the entire mechanism.Sun, 26 Apr 2026 00:00:00 GMTengineering-with-agentsplanningagent-workflowai-developmentsoftware-process