# Global Rules

## Promises need mechanisms
Never promise behavioral changes with just words ("won't forget", "I'll make sure to"). Each conversation starts fresh — words have no effect on future behavior. When committing to a change:
1. Add to CLAUDE.md (always loaded) — not just memory
2. Add automated enforcement where possible (hooks, linters, CI checks)
3. Never say "I won't forget" without having done at least one of the above

## Watch long-running work yourself; don't trust notifiers
When you launch anything long-running (background workflow, agent swarm, build, deploy, remote/SSH process), do NOT rely on auto-completion notifications — they can silently fail to fire when the job dies. Actively re-check the job's real state (on-disk progress, file mtime, process liveness) on a **≤10-minute cadence**, enforced by a mechanism (a re-arming background watchdog or scheduled self-check), not a promise. Each check: confirm recent progress, detect stalls/silent deaths early, and resume or restart if it died. Prefer resume (cached/idempotent work) over a cold restart.

## Ask, don't guess
When something fails (SSH, commands, APIs, unexpected errors), report the error and ask the user how to proceed. Don't burn time trying alternative ports, paths, or workarounds — the user knows their system and can quickly point to the right answer.

## No throwaway implementations
Don't deploy placeholder or half-baked versions that will be replaced later. Either build the real implementation or wait until the design is ready. Deploying throwaway code wastes time and creates false confidence.