Prompts to Pipelines v1: Best Practices for Claude Code
Self-contained HTML deck. Arrow keys, click, or swipe to navigate. A theme toggle lives in the top-right corner of the deck.
Abstract
This is the first cut of the Claude Code harness talk, given two weeks before the iterated v1.5. The bones are the same: five plagues of agentic development, a harness-based development model, the primitive map. What is different is the emphasis. v1 spends more time diagnosing the problem space and more time on the community frameworks that were emerging at the time, and uses the rdf pipeline as a worked case study rather than fixtures.
If you are new to the material, v1.5 is the version to read. This one is here because the evolution between the two is itself useful. Six months of running the harness in production surfaced things the first version got wrong and a small number of things it got exactly right. The diff between v1 and v1.5 is a decent proxy for what matters and what does not.
Key Takeaways
- The five plagues are real and named the same in both versions. Once you can point at the failure mode, the fix tends to be obvious.
- Community frameworks come and go. The primitives are stable: memory, hooks, subagents, settings.
- rdf started as a three-file convention and grew into a pipeline. The growth was driven by specific incidents, not design.
- The CLAUDE.md in v1 is shorter than the CLAUDE.md in v1.5. Not because the project got more complex, but because the rules kept landing.