I'm a developer and AI enthusiast. I share my journey and thoughts about AI and technology here.
What every agent should follow everywhere goes global; what only matters to one project stays in that project. A rule gets added when a problem keeps coming back, a skill when a task really needs finer guidance.
I use AI every day and still have nothing to show for it. Between getting the job done and building a product there is a gap, and I think that gap is sensitivity to pain points.
The skill name says it all: BDD to define how the product should behave, handoff to turn the discussion into documents someone can pick up, and a lead agent with a few sub-agents to get the work done.
Setting up a personal wiki takes five minutes; maintaining it has no end. Keep originals, summaries, and topics apart, keep one entry point for rules, isolate drafts, and check content as well as files.
After reading the claude-code source, I found it does not rely on a simple system prompt to nudge the model toward correct tool selection. Instead, it uses a layered control surface to shape behavior in advance.
After reading the claude-code source code, I found that its tool calling optimization relies not on a single trick, but on multiple layers of mechanisms working together.
I prefer agent memory on demand, not by default. The real difficulty is deciding what to keep, when to keep it, and how to stop memory from turning into noise.
Claude Code is great for interactive exploration. Once you need long-running, recoverable, auditable agent execution, code-level control becomes much harder to avoid.
Yes. It is git worktree used as isolation for agent workflows: one repository, multiple directories, and far fewer workspace collisions.
Prompt caching is driven by prefix matching, not by session identity. If the same prefix shows up again, reuse can happen across conversations and sometimes across users.
TTL is the lifetime of a prompt cache entry. Each hit refreshes it. Leave it unused for long enough, and it expires.
Prompt caching cuts the price of repeated prefixes. Tool-result replacement shrinks the prompt itself. Both can save money, but they push on different parts of the bill.