A new paradigm for Claude Code Agent Skills whose steps actually depend on each other. The current pattern - what every Skill on the market does- is dump a long 
SKILL.md into the agent’s context and hope it follows the instructions in order. It mostly doesn’t: the agent skips prereqs, hallucinates outputs that “look right,” can’t resume after /compact, and produces a different run every time you ask. ClawBio 2.0 replaces the hope with a contract: each skill ships a 
pipeline.yaml DAG, and a ~500-line CLI exposes start → next → open → done → bundle as the only legal moves. Instead of trying to read prose & remember, the agent calls CLI next, gets told exactly which step is unblocked, and is physically refused if it tries to skip ahead. State lives in one atomic 
state.json that survives /compact and /clear. Live I’ll show: starting a multi-step run, the CLI rejecting an out-of-order step, the agent recovering cleanly from a forced /clear, and the reproducibility bundle (
commands.sh + checksums + env) that a Stop hook refuses to let the agent skip. Bioinformatics is just the testbed - the pattern is for any skill where order matters and outputs feed inputs.
