Claude Code
Reviewing what Claude Code changed
Claude Code can change dozens of files in one session. How to review its work properly: see each decision on the lines, push back where you disagree, and keep the judgement yours.
Claude Code is fast. One session can touch thirty files, and the summary at the end names a dozen of them with line numbers. Reviewing that properly is the hard part.
The problem with the end-of-session summary
The summary is accurate, and it is still hard to use. Each file:line is something you have to open, and the argument only exists once you have opened all of them and held them together in your head. Most people stop after the second file and type "looks good".
Ask for the review as a deck
deck gives Claude Code another way to answer. Instead of a paragraph, it writes a short deck: one claim per group, with the code that proves it. A bar appears at the bottom of your screen, and you open it when you are ready.
Inside:
- each sentence lights the lines it is about
- a proposed change shows as the old lines going and the new lines arriving
- behaviour over time, like a queue backing up, plays as a chart in step with the words
- you select lines and comment, or ask, and the same session answers
When you submit, the command Claude Code ran ends and hands your review back, pinned to the lines. It continues with your decision in hand, in the same session with the same context.
Setting it up
Run the installer, then deck setup. Setup installs a short skill so Claude Code knows when to reach for a deck, and it offers the catch: a hook that notices when a reply names two or more file:line locations and asks for a deck instead. After a long session is compacted, the same hook reminds it to load the skill again, so decks late in a session are as good as early ones.
What to ask for
- "Show me what you changed, one decision at a time."
- "Before you write it, show me the plan."
- "Why does this break only under load? Show me."
The answer comes back as something you can check, not something you have to believe.
Questions
Does deck run its own model?
No. Claude Code writes the deck with a few shell commands and puts it on your screen. deck draws it and hands your answer back to the same session.
Is the catch required?
No. It is offered during setup and can be turned off. Without it, Claude Code still uses deck when the skill suggests it.
Does it work with other agents too?
Yes. Anything that can run a shell command can write a deck, including Codex, Cursor and Amp.