ai notes.
Every note in one place. Concepts I wanted to explain to myself, lessons from building with agents, and the odd line from the log.
- ConceptCoding Patterns Stop AI ForgetfulnessUsing a strict coding pattern helps an AI write better code by giving its limited short-term memory a clear blueprint to follow.
- ConceptAI + Good Architecture = Quality CodeAI doesn't change the fundamentals of good software — it amplifies whatever foundation you give it.
- ConceptBuilding software products with AIMindset shifts and practical approaches for building software products with AI assistance.
- ConceptDesigning Agentic WorkflowsThe fundamentals of designing, evaluating, and debugging agentic AI workflows.
- ConceptDesigning Agents as Teammates Not ToolsEffective agentic AI requires treating agents as teammates who need proper context and experience design, not as tools you configure with the right parameters.
- LessonSoftware engineers won't need to write and read code only when agents will embed good engineering practices in their DNAAs long as we have to tweak the code ourselves, I don't think much will change. In fact, good engineering practices help AI generate better solutions.
- LessonThere's not difference in maintainability in AI vs Human generated codeLLMs are great at pattern recognition, so as long as the codebase is well designed and architected, it'll produce high-quality output.
- ConceptWe should invest in AX (Agents Experience)AI agents are becoming our new teammates: we should treat them as a new entity that requires its own UX to work properly, a bit like DX.
- LogPlanning (Agentic Pattern) with code improves performanceLetting the LLM express its plan in software code that you can just execute can be a very powerful way to let it write rich plans.
- ConceptAgents can be evaluated objectively or subjectivelyCan the evaluation condition be expressed mathematically? If so, we can evaluate with code. If it can't, it's a task that requires non-deterministic thinking.
- ConceptAgentic workflows are only as good as their weakest component + error analysisAgentic workflows follow the same principle as Theory of Constraints: look at the traces, count the errors, fix the bottleneck.
- Concept"LLM as a judge" grading with a rubric gives more consistent results when evaluating Reflection (Agentic pattern)"LLM as a judge" can help by providing a rubric that the LLM can use in the reflection prompt to judge the output.
- LogReflection (Agentic Pattern) can use external feedback from tools to improve outputBy using a tool for web research, the LLM can adjust course with the external information.
- ConceptAgentic design patternsReflection, tool use, planning and multi-agentic workflows.
- ConceptAgentic AI can be less or more autonomousFrom all steps predetermined to agents that make many decisions autonomously and can create new tools on the fly.
- ConceptAgentic AI works better for well know, linear processesAgents work better with a clear, step-by-step process and standard procedures to follow. Its job gets harder if the steps are not known ahead of time.
- ConceptSoftware systems integrating with AI still benefits from good software architectureAI is infrastructure, not architecture. The good practice is still to create an abstraction exposing a stable API.
- LessonCoding with LLMs makes me think moreWhen working with AI-assisted tools like Cursor, I'm forced to think more about what I want the agent to do, so that it doesn't get lost.
- ConceptContext compression for better code generationResearch, planning, implementation: an iterative process aimed to improve the AI context.