Specify before building (L0 to L5)
Gates: L0-understand, L1-size, L2-decide, L3-contracts, L5-freeze. Each gate ends with the verify skill.
- L0-understand: write the problem statement, users, constraints and success measures into the spec, with persisted research that cites its sources and marks what is UNVERIFIED.
- L1-size: sizing.md lists every component with owner role, estimate and risk:
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" size lint - L2-decide: one ADR per decision, then the technology matrix (pinned version, license, evidence for each choice):
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" adr new --title "TITLE" --applypython "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" adr lintpython "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" tech lint - L3-contracts: a data contract for every interface, listed in the catalog, then the parallel waves:
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" contract lintpython "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" contract graph - L5-freeze, the spec (EARS requirements, each with a runnable Gate; interfaces; tasks with waves):
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" spec new --title "TITLE" --tier M --applypython "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" spec lint - After tests exist, build the trace matrix:
python "${CLAUDE_PLUGIN_ROOT}/scripts/batuta.py" spec trace SPEC_ID --apply - When lint is clean, ask the human to run
/batuta:freeze(only a person can start it). You never approve a spec yourself.
On a non-zero exit: exit 1 is a refusal (a requirement without a Gate, an id reused, a missing ADR for tier L, an open
[NEEDS CLARIFICATION]): fix the artifact and run the same command again. Exit 2 is a warning (for example a component
without a unit): resolve it or state it in the gate's evidence.