Adam Lewis

Team Rules outrank project rules, which outrank user rules, but what happens when a guardrail in the ledger contradicts a rule someone committed, I cannot tell you. Does brief just hand it over, or does it flag the clash and let a human settle it? Whether the ledger is a fourth layer or a second source of truth depends on the answer.

Detail the metrics list. Revision cycles and launch time are easy to count but they don't tell you whether the output was any good. A workflow that halves the cycles by approving worse drafts will score well on every one of those measures. So what's the quality signal sitting alongside them, something a reviewer marks, a rejection rate after publication, or just the human-oversight step the article keeps mentioning. Without that the numbers only prove the process got faster. https://tone-of-voice-generator.com/existing-brand-guidelines-into-ai

Denying a full read of a lockfile is cheap to get right, but a hook that guesses at size will wave through a 20 KB generated file as often as it blocks a bad one. Deny rules are blunt too, at least they're predictable. If the hook gets it wrong, the model burns a turn arguing with it before doing the grep you wanted.

A 32 KiB file is big enough that the rule you care most about often falls off the end. Is there a usual ordering that survived in the files you looked at, or do teams just append and let the tail take the hit? The hook-over-prose point is well made either way, since a permission deny is read by the harness.

The Stop hook keeps the harness honest, because everything before it only stops the agent doing the wrong thing while it's still running. Someone writes a rule down and later edits it out. The Stop hook won't catch that. Rules that people amend by hand need to live somewhere the generator doesn't touch. I've not found another way to close that gap.

Reviewers read the diff because that's the artefact they've got. Grep the changed symbol before you trust the change, since the diff will look exactly as clean either way. Signing off on an agent's patch means signing off on that search.

Walk through the gotchas properly. Sharp surviving a Next.js bump and the advisories npm audit misses both read like things found by getting it wrong first. Did any of those turn up after you'd started trusting the checker's exit code?

A context file is useless while the agent is editing the one file the rule is about. What changed behaviour for me was moving the boundary into the build so a bad import fails the run before anything gets committed. Types catch the signature. Lint catches the direction of the dependency.

Naming the states is what takes the time. If you can't say what state the system is in when the thing goes wrong, you don't have a failure case, you have a try/catch. I think that is the job. It feels finished because the error handling is there.

Review is being treated as a first-class problem. I care about what happens when a comment pins to a passage the agent then deletes outright, since a diff that anchors to old text has nothing to anchor to once that text is gone. Does the tooling hold the comment somewhere until the passage comes back, or does it just drop it? https://prickles.org/tenet/the-intern-pattern/AI1