What actually fails remotely
Distance gets blamed for problems that are really ownership and visibility gaps.
| Symptom | Usually means | Fix |
|---|---|---|
| Surprise scope | No written decisions | Short ADRs + shared backlog |
| “Still in progress” | No demo cadence | Working software every 1–2 weeks |
| Model mystery | Choices live in chat | Prompt/tool notes in the repo |
| Blame across zones | No named lead | One delivery owner both sides |
Named owners beat ticket ping-pong
One delivery lead, written decisions, and a shared backlog keep timezone gaps from becoming blame gaps.
- Agree overlap hours for decisions—not for watching status.
- Put acceptance criteria where the work lives, not only in email.
- Escalate blockers to people, not to a faceless queue.
Demo on a cadence
Working software every week or two beats slide updates. For AI features, show eval samples and failure modes, not only happy paths.
- Book a fixed demo slot across timezones.
- Show the slice running—happy path and one failure.
- Review eval samples or logs when ML is in play.
- Capture decisions and next acceptance criteria in writing.
Document the “why”
Model choices, prompt/tool boundaries, and rollout flags belong in the repo or a short ADR—so the next engineer isn’t guessing.
