Hey Reader,

Short answer to the subject line: yes – but only with the workflow underneath. "Unsupervised" doesn't mean nobody's reviewing, it means the human moves from typing to reviewing. The three reads below are the layers of that workflow stack.

Plan Mode is how you organize a session before the agent runs. Source control is the safety net that makes letting Claude Code ship code on your behalf safe instead of reckless. 2.1.141 is the runtime hardening that keeps your work alive across reopens. Three layers, one stack.

How I shipped a 3-week-old design handoff in 18 min (Episode 2)

Episode 2 of the Production Claude Code series is up. I had a design handoff sitting in a folder for three weeks – the kind that turns into hours of "where does this go again?" once you finally open it. Plan Mode shipped it in 18 minutes.

Plan Mode is the first of Claude Code's 4 execution modes. The model writes a structured plan – tasks, files, ordering – without touching your code. You read it, edit it, approve it. Then Auto Mode executes the approved plan without interrupting on every tool call. That separation is the unlock. Planning is cheap. Execution is mechanical.

The post walks through the actual session: the prompt I used, the plan Claude wrote, what I changed before approving, and the 18 minutes of execution that closed the handoff.

A beginner's guide to source control with GitHub

People keep DM'ing me asking how to use source control. Not "what is Git" exactly. More like "I'm building a thing with Claude Code and everyone says I need GitHub, what does that actually mean and what do I do first." So I wrote out the answer once.

The post covers the five Git commands that cover almost everything, branches and pull requests, conventional commits, and the two GitHub-specific advantages most builders underuse once Claude Code is in the loop – the `gh` CLI as the cleanest agent integration surface, and Copilot's automatic PR code reviews stacking a second AI reviewer on top.

Personal admission in there: these days I let Claude Code handle the commits, the pulls, the pushes, and the PRs. I just review the PRs. That only works because the branching discipline and the PR review checkpoint exist underneath.

Claude Code 2.1.141 – the stabilization pass after 2.1.139's feature wave

Eight releases since the last recap (2.1.131 → 2.1.141), and the last one is the most useful one. Roughly 60 fixes, plus four real adds most readers will miss if they only scan the version number.

The four that matter:

→ `terminalSequence` in hook JSON output – desktop notifications without owning the terminal (critical for IDE-hosted sessions).
→ `ANTHROPIC_WORKSPACE_ID` – workload identity federation finally gets workspace-level token scoping.
→ `claude agents --cwd

Plus the polish wave: background agents stop reverting permission mode on reopen, `/model` stops mutating autocompact thresholds across sessions, "view diff in your IDE" is restored on file-edit prompts, and the silent 1-hour prompt-cache bug from 2.1.131 is finally fixed (your prompt cache TTL is back to 1 hour, not 5 minutes).

If you skipped the 2.1.131 changelog because the version number didn't look special, this recap brings you current.

Reading order if you want the through-line: source control first (the floor), then 2.1.141 (the runtime), then Episode 2 (the workflow that runs on top).

Reply and tell me which one moved a needle for you. I read every reply.

– Alex

P.S. – Last day for early bird on the 5-day workshop. Production-Grade Claude Code in 5 Days kicks off June 1. Early bird is $297 through 11:59 PM PT tonight (Friday May 15). Tomorrow it goes to $497.

Same setup I use to ship daily, unpacked over 5 days – plan mode, auto mode, hooks, MCP, the agents view, AGENTS.md, source-control conventions, and the full Compound Engineering pipeline on a real production codebase.

Save my seat (early bird ends tonight) →