The Loop Runs While You Read. Plan Mode Is The Pause.
This is Liam, in for Bear. One sentence. Watch a chatbot cannot do this. Two Reads. A finding. Two Edits. A Bash run. Seven tool calls, two files changed, and tests pass — in one turn. Thirty-eight seconds. Nobody said yes to any of it.
This is Liam, in for Bear. One sentence. Watch a chatbot cannot do this. Two Reads. A finding. Two Edits. A Bash run. Seven tool calls, two files changed, and tests pass — in one turn. Thirty-eight seconds. Nobody said yes to any of it.
Series: Claude Code 101 Topic: CLAUDE CODE 101
CHAPTERS
0:00 Introduction 0:31 Definitions 2:50 Verdict 3:16 Your Turn
YOUR TURN
Your turn. Before your next non-trivial edit, press Shift+Tab twice to enter plan mode, and type your ask. Read what it proposes. Then approve, or edit the plan first. Once. Notice how much of the loop you would have missed.
@NikBearBrown / Liam-in-for-Bear
Part of the Claude Code 101 series.
Sources and citations: see the reel's SOURCES.md in the open textbook.
What this video covers
- 0:00 — Introduction
- 0:31 — Definitions
- 2:50 — Verdict
- 3:16 — Your Turn
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
Introduction
[0:00] This is lay in forbear. One sentence. Watch. A chatbot cannot do this. Two reads, a finding, two edits, a bash run. Seven tool calls, two files changed, and tests pass in one turn. 38 seconds. Nobody said yes to any of it. Here's the idea of this film. Claude Code isn't a faster chatbot. A chatbot produces text and waits. Claude Code runs a loop. Gather, act, verify, and closes it on its own until it thinks it's done. Plan mode is the one place the loop stops. Same request, same tools, same repo, one flag is the difference. Four words
Definitions
[0:31] Before we start. Agentic loop gather, act, verify, and repeat. Claude codes own cycle run without you between steps. Tool call one atomic thing. Claude does read a file, run a shell, edit a line. Permission mode, the flag that decides which tools may fire without asking except edits, fires them. Plan fires nothing. Exit plan mode, the interruption point, the tool call that puts the plan in front of you and waits for yes. Check the loop run. Get diff. Five lines added, two removed, two files touched mine. Now whether I read them or not, the tests pass four of them. And the count seven tool calls one turn 38 seconds. Claude's check mark and my count agree. That is not the question. The question is when I got to look reset, same folder, same files, same ask, one flag change, not a better prompt, not a smaller model, not fewer allowed tools. The same seven tools are on the allow
List. Only the permission mode changes except edits becomes plan. Watch the loop's shape change with one flag. Same sentence. Plan mode. Two reads same files. A finding the same bug. Then something except edits. Never does. It writes the plan to a file. Then a tool call I've never seen before. Exit plan mode. Carrying the plan and the shell commands it would like. Next, and it stops. Exit plan mode. In headless, there's nobody to answer in the app. That is your keystroke. Check the plan. Run. Get status. Nothing to commit. Scratches untouched. The plan is on disk. 28 lines at a path Claude chose with context change and verify and a test span. Same ask zero writes to my repo. The loop stopped at gather. Who did what? Step one mine. One sentence two permission modes. The whole experiment. Claude did the loop. Seven tools in one turn. Handoff was tests pass without a pause and it met it. Step
Three the dangerous middle. Two files were mine before I could read either. Then me again. Reset and add one flag. Claw did the plan run. Three reads one right. then exit plan mode and stopped. Then interpretive judgment. The same loop can close on its own or wait for a keystroke. I choose which. So what was mine? Claude can run seven tools in one turn and it will. It can close the loop without pausing and it did. It should propose a plan before writes. It does. When I ask it should wait on exit plan mode and it does. I must choose the permission mode. Nobody else can pick it for a session on my repo. I must read the diff, not the summary. The summary is Claude's. The diff is mine. I must answer exit plan mode myself and I should reach for plan mode first on anything I would not want undone by hand. Let's recap with Claude. One
Verdict
[2:51] Sentence in accept edits. Seven tool calls, two files changed. Tests pass in one turn. No pause. Same sentence in plan mode. Three reads one plan files. Zero writes to my repo and the loop stops on exit plan mode. The loop is one design. Plan mode is the interruption point that lets a human read before act. What would prove this real wrong? A plan mode run that edits a file in Scratch without a human answering. Exit plan mode first. Your turn. Before your next
Your Turn
[3:17] Non-trivial edit, press shift plus tab twice to enter plan mode and type your ask. Read what it proposes. Then approve or edit the plan first once. Notice how much of the loop you would have missed. The loop runs while you read. Plan mode is the pause. Lay them in for bear.
