Compound-engineering, in practice
Every project at Crewtives starts with a KICKOFF.md and ends with skills you can run, not docs you have to read. The test is whether someone — human or agent — can resume your work without you in the room.
The pitch for compound-engineering is that every action a human takes should be takeable by an agent. The real test is narrower: can someone resume your work without you in the room? If the answer involves “let me DM the person who knows,” the system isn’t compounding yet.
KICKOFF.md is the contract
Before any code, we write KICKOFF.md: what we’re building, what we
explicitly aren’t, the copy that’s locked, the decisions we won’t
re-litigate. By week two — when scope creep starts asking polite
questions — this is the doc PR comments link back to. It’s also the
first file the agent reads when picking up a thread it didn’t start.
Skills, not docs
When we solve a problem worth keeping, we write a skill instead of a
runbook. The difference is who exercises it. A runbook gets read once,
rots, and gets re-read incorrectly under pressure. A skill gets invoked
by /skill-name and runs the same path every time. Our “commit, push,
open PR” flow used to be a five-step doc nobody opened. It’s now one
invocation that runs the same checks, writes the PR body in the right
shape, and never forgets the pre-push step.
The corollary: docs that didn’t survive the translation to skills mostly turned out to be docs nobody actually needed.
Plans before code, always
Every non-trivial feature begins with a plan doc: outcome, milestones, verification. The plan is what lets multiple workers — agents and humans — pick up pieces in parallel without colliding. Without it, two tracks race for the same file, decide differently, and you spend a day reconciling instead of shipping.