baton-planlisted
Install: claude install-skill StanimirTenev/baton
# Baton plan — goal → research rounds → plan
A big task fails in two quiet ways. It is planned from what the agent already believes,
or it is researched with no question in mind. This skill runs a loop that avoids both:
**goal → questions → research round → verified facts → plan vN → human decides → next round**
It stops when no open question still blocks the plan, or after **2 rounds**. Two rounds
are the default because the second already changes the plan substantially and a third
returns less than it costs; run a third only if the human asks for it. What is still
unclear by then goes to the human as a decision, or into the plan as a task (a cheap
test, a question for a lawyer). After that the plan is **executed**, not researched further.
Task root and logbook name: the session-start board names the task root in its first line
("Baton — the tasks in …"); installed as a plugin, the rules that come with it name both. With
the install script they are also in `~/.claude/baton/hooks/baton.local.json` (`home`,
`logbook`; defaults `~/tasks`, `LOGBOOK.md`). The files below are written into `<task root>/<task>/`.
## Before you run this at all: is it a swarm question?
Baton has two paths, and this skill is the expensive one. Most work is not this.
**The direct path** is what the hooks already enforce and needs no skill: read the task's
logbook, do the work, write the entry. One letter, one fix, one decision, one measurement you
can take yourself. It costs what the work costs.