For Anthropic's Claude
Stuck in Claude? Press the dot.
Claude is excellent at the gnarly parts. The press is for when 'excellent' isn't quite enough.
When Claude needs a person.
01
When the architecture decision is too consequential to leave to a chat
Claude has good intuition; for choices that lock in cost and complexity for the next 18 months — multi-tenant data model, async-job runtime, auth strategy — the press is for a software engineer who's lived with the tradeoffs.
02
When the refactor exceeds Claude's effective context
Even a million-token window has limits when the refactor is across 200 files. A Relay engineer breaks it into reviewable pieces, runs each through Claude, and verifies in between.
03
When you need a person on the deploy
Claude can write the IaC; deploys still go wrong. The press covers the run.
Common questions
- Does Relay use Claude internally?
- Our co-pilot for engineers uses Claude Haiku for risk-scoring and session summarization. Customer code is never used to train the model.
- Can the engineer pair with me in Claude.ai?
- Yes via screen share. The engineer reviews the conversation, suggests reframes, and helps you turn the output into shipped work.
- What about Claude Code (the CLI tool)?
- Fully supported. A common press: hand off a Claude Code session to an engineer mid-build for the steps that need a human in the loop.
Related reading
- Field notes · 9 min readAnatomy of the handoff: what we learned shipping the first 1,000 sessionsA median match in seconds isn’t a marketing number; it’s a system constraint we engineered toward and missed by 16 seconds in the first month.
- Research · 17 min readWhen does an AI build want a person? A study of 4,200 stuck momentsWe logged every press in the first quarter of the private beta. Four patterns explain almost everything.
- Essay · 11 min readThe irreducibly human moment in softwareAn argument for why software engineering doesn’t shrink in the age of AI, it sharpens, and moves to a new place in the build.