Lay down the three-file minimum workbench (router, state, task board) into a fresh workdir/ and prove a single agent turn can read state, pull a task, write to scope, and persist updated state.
- An empty
workdir/directory next to the lesson code - Knowledge of the three files:
AGENTS.md,agent_state.json,task_board.json
code/main.pythat creates the three files and runs one turnworkdir/AGENTS.mdshort router pointing at state, board, and the verification commandworkdir/agent_state.jsonwith active task id, touched files, next actionworkdir/task_board.jsonwith a small backlog and statuses
python3 code/main.pyexits zero on first and second run- Second run picks up where the first left off, not from scratch
- Diff printed by the script shows the one file the turn touched
- Scope contracts, verification gates, reviewer agents. Those layer on top in later lessons.
- Long monolithic
AGENTS.md. The router stays short on purpose.
docs/en.md- full lessoncode/main.py- reference implementationoutputs/skill-minimal-workbench.md- extracted skill