Welcome to my little workshop
Building Hope-Driven Products
Writing about AI-native building, games, open source, and trade.
Explore the field notesFresh from the garden
Latest field notes
-
When the Model Eats the Harness
As models absorb common agent skills, simpler harnesses become easier for users while moving more of the product advantage into the model itself.
-
agentsnippet: Reuse AGENTS.md and CLAUDE.md instructions
agentsnippet turns reusable Markdown snippets into ordinary AGENTS.md and CLAUDE.md files, keeping the instructions that invoke coding-agent skills consistent across projects.
-
Short Automation Prompts, Long-Lived Skills
For automations I expect to maintain, I prefer a short prompt that invokes a versioned skill instead of pasting a long instruction block into the scheduler.
-
Loop Engineering Reduced to Two Patterns
Loop engineering becomes easier to reason about when most workflows are reduced to two shapes: filter many candidates down, or improve one artifact until it passes.
-
Human-in-the-Loop, Not Human-as-the-Loop
Planrock gives coding agents durable plan memory, while /goal keeps execution moving so humans can supervise judgment instead of manually restarting every step.
-
AI Plan Mode Is Good. Saved Plans Make It Better.
Plan mode helps agents slow down before coding, but saved repo-local plans make that structure durable across sessions, machines, and agents.
-
A Two-Layer Project Structure for AI Coding Agents
Keeping code in focused repositories while using a higher-level workspace as the operating layer for AI agent memory, planning, and cross-repo work.