Warp Speed Systems — Warp 2: Pair Worker

You're in the room. Use that.

Same input → model → response loop as Warp 1 — but the model now edits your codebase, runs commands, opens PRs. A bad answer at Warp 1 costs nothing; a bad commit costs a rollback. The new lever: a persistent instruction file that survives across sessions and turns repeated corrections into rules the model already knows.

The short leash

Budget: $20/month. Claude Code in your terminal or IDE. You stay in session, watch what Claude does, and review every change before it ships. Quality is bounded by your attention — so is speed. At this warp, the bottleneck is the quality mechanism.

The persistent instruction layer

At Warp 1, every session starts blank. At Warp 2, a file named CLAUDE.md in your project gets loaded into the prompt every session, automatically. It's the first layer the model reads, on top of whatever you type that day.

The loop: review the response, catch a recurring mistake, write the correction into CLAUDE.md. Next session starts knowing it. If you have to say something twice, it belongs in the file.

What automation works at this warp

Some autopilot is safe — recoverable, narrow scope:

• Auto-merge to dev — never to main. Dev is where you review; main is production.
• Issue tracking — creating and updating tickets as work progresses. Low risk.
• Narrow subagent tasks — isolated work with specific inputs and outputs. You check the result.

Writing good rules

“Production quality” is not a rule. “No console.logs in committed code; error handling on every external API call; types on every function signature” is a rule. A rule is good when applying it doesn't require judgment.

Every correction you make is a clue. Why did the model produce that? Almost always: your prompt left the standard implicit. Write what you wished the prompt had said. Over weeks the file becomes your actual standards in a form the model can read — not the version in your head, the version that survives contact with another reader.

The bottleneck is you — on purpose

Being present forces you to understand the work, not just approve it. The goal isn't maximum throughput; it's producing quality reliably before producing it quickly.

At some point review starts feeling redundant. You read a PR and already know it's good because the standards in CLAUDE.md made the output predictable. That feeling is graduation approaching.

The graduation condition

You're ready for Warp 3 when your standards live in writing, not just in your head — when you could hand CLAUDE.md plus a short brief to a worker that isn't youand trust the result — because if it came back wrong, you'd know the brief was what failed. That shift — from “Claude did something wrong” to “my instructions were insufficient” — is what lets you stop pairing with one agent and start briefing a team.