One Baseline for Every AI Coding Tool
Wilkes & Liberty runs its own delivery work through AI coding agents — several of them, from different vendors, on every engineer's machine. That practice surfaced a problem most teams adopting these tools will meet within a quarter: each tool keeps its own configuration, in its own format, in its own corner of the home directory. Skills lived in one place for one tool and another place for the next. Server connections were wired by hand. When something broke, it broke on one machine, silently, and stayed broken until someone noticed the tool had quietly stopped doing part of its job.
The problem, concretely
Three symptoms kept recurring. First, drift: two machines that should have behaved identically did not, because one had been configured by hand months earlier. Second, silent breakage: a tool would lose a capability — a skill directory unlinked, a server connection dropped — and report nothing, because nothing had crashed. Third, unpaid maintenance: every improvement to one tool's setup had to be repeated by hand for the others or it simply never happened.
The approach
The fix was to treat AI tool configuration the way we already treat infrastructure: as versioned code with one source of truth. A single baseline repository now declares everything the tools share — operating rules, skills, agent definitions, server connections, and pinned third-party skill collections with their licenses recorded and checked. Thin per-tool adapters translate the one baseline into each tool's native format. An activation command applies it; a doctor command verifies that what is live on the machine matches what the baseline declares, and warns when it does not.
Two design rules did the most work. Policy must be enforced where it is written — a license requirement that the sync code does not check is a wish, not a gate. And every check must be able to fail — a health command that only reads its own ledger will pass forever while the world rots around it.
Results
An engineer sets up a new machine by cloning one repository and running one command. A change to any shared asset — a new skill, an upgraded third-party collection, a new server connection — ships as one reviewed pull request and lands on every machine and every tool through a pull. Drift now announces itself instead of hiding: the doctor command found and named configuration failures that had previously gone unreported for hours, including one where a tool reported success while skipping the work.
The baseline is also how we evaluate what we recommend to clients. We run the reference implementation of the practice before we advise anyone else to adopt it.