Skip to content
Getting started·5 min read·English

The decisions Claude Code makes on its own mid-task

In a long task, Claude Code makes small decisions on its own: who counts as the same customer, what goes on the list, what is allowed. Most of them are fine. Some you only see if you read every line. You should know is a side agent that ships inside Claude Code, reads the work as it happens, and writes a note above the prompt when there is something worth knowing. It is off by default. In this guide: turning it on, what a note looks like, and three questions for the end of every long task.

The decisions Claude Code makes on its own mid-task, and how to turn on You should know

Who this is for

  • You give Claude Code tasks that run for minutes and you do not read every line? Start with part 1, turning it on.
  • Only want the questions? Jump to part 5. They work without the plugin too.
  • Claude Code not installed yet? Start with the day one guide and come back here.

Contents

  1. 01.Turning it on, step by step
  2. 02.What a note looks like
  3. 03.The example from the Reel
  4. 04.Why it is quiet most of the time, and what it costs
  5. 05.Three questions: what Claude decided on its own
  6. 06.Turning it off
  7. 07.Sources
PART 01

Turning it on, step by step

The simplest way is from the terminal, the window where you start Claude Code. If Claude Code is open, leave it first with /exit. Paste this and press Enter:

Copy
claude plugin enable cc-plugin-you-should-know@builtin --scope user

You should see this line:

✔ Successfully enabled plugin: cc-plugin-you-should-know

Then open Claude Code as usual, with claude. The terminal command turns it on for every project on the computer, not only the current folder.

Already inside Claude Code?

You can type the command Claude Code itself suggests in its tip:

Copy
/plugin enable cc-plugin-you-should-know@builtin

How to tell it worked

  • In the terminal: the line with ✔ Successfully enabled plugin appeared.
  • In a long task: at some point a line starting with ✦ appears above the prompt. Not in every task, and that is on purpose (part 4).

Did not work? In this order:

  1. It says command not found: claude? Claude Code is not installed in this terminal. Start with the day one guide.
  2. It says the plugin was not found? Update Claude Code with claude update and try again.
  3. The command worked but no notes show up? Short tasks almost never get one. Try a task of a few minutes with several steps.
  4. You run Claude Code from a script with claude -p? It does not show up there at all, because there is no screen to show the note on.
PART 02

What a note looks like

The note appears above the prompt with one of two tags:

  • Heads up: about the work in this session. A decision Claude Code made, something it did not highlight, or a result that may be off, when there is a cost if you miss it.
  • You should know: how something works, when understanding it matters for your work.

Under the note there are buttons. Press the number on your keyboard:

  • 1: Learn more opens a short explainer of a few lines.
  • 2: Knew this already tells it you knew.
  • 0: Dismiss closes the note.

After Learn more the buttons change: 1: Understood, 2: Chat in main session, which moves the topic into the main conversation, and 0: Dismiss.

PART 03

The example from the Reel

In a demo folder with made-up customers I asked Claude Code for one list for a WhatsApp campaign, from three files: a shop export, WhatsApp contacts and a newsletter list. I asked it to work in small steps: read each file, write a script, run it, write a test, and commit.

Along the way it set a rule on its own: anyone who signed up for the email newsletter goes on the WhatsApp list. Four of eight people were left.

At step six, 12 seconds before the list file even existed, this note appeared. This is the original output:

✦ Heads up · The campaign list will drop about half the customers, and even the people kept haven't explicitly agreed to WhatsApp marketing. 1: Learn more 2: Knew this already 0: Dismiss

And after Learn more:

WhatsApp list uses email opt-in as consent - None of the three files records whether a person agreed to get WhatsApp marketing messages. - The main agent plans to treat newsletter sign-up (email consent) as permission, and to keep only people who have an active WhatsApp contact. - That leaves about 4 people: Dana, Yossi, Avi, Maya. It drops Noa (blocked), Tal (unsubscribed), Lior (no WhatsApp contact) and Shira (her records only match by name). - In Israel, marketing messages usually need clear prior consent, and opting in to emails may not cover WhatsApp. I'm not fully sure how the law treats this case. - Before sending, check that these 4 people actually agreed to WhatsApp messages. 1: Understood 2: Chat in main session 0: Dismiss

The note also has a line about the law, with the side agent's own hedge. That is its wording, not legal advice.

The agent doing the work said the same thing only at the end, in one line at the bottom of its summary:

Decision for you: newsletter consent is consent to email, not explicitly to WhatsApp. If you have a WhatsApp opt-in record, that's the right thing to filter on before you send.

That is the whole difference: the note came in the middle of the work, while you can still stop it. The line in the summary came after the list was already done.

PART 04

Why it is quiet most of the time, and what it costs

  • Its instructions say the bar for showing a note is very high, and when in doubt, show nothing.
  • It only checks during a long task, every few steps. A 30-second task will almost never get a note.
  • It skips topics you already talked about in the main conversation.
  • Notes you ignore come back less: after a few, it checks less often.
  • Every check runs the model, so it counts toward your plan's usage.
  • In claude -p, the screenless mode you run from scripts, it does not show up.
PART 05

Three questions that get Claude to tell you what it decided on its own

The plugin does not catch everything, and it is not built to. So at the end of a long task, before you use the result, paste one of these:

1. Its decisions

Copy
Which decisions did you make on your own in this task that I didn't explicitly ask for?

2. What went missing on the way

Copy
What did you delete, merge or skip, and why?

3. What to check with your own eyes

Copy
What is the one thing I should check with my own eyes before I use the result?
PART 06

Turning it off

From the terminal:

Copy
claude plugin disable cc-plugin-you-should-know@builtin --scope user

You should see:

✔ Successfully disabled plugin: cc-plugin-you-should-know

Or from inside Claude Code:

Copy
/plugin disable cc-plugin-you-should-know@builtin
PART 07

Sources

  • The plugin's description and the side agent's instructions were read from Claude Code itself, versions 2.1.323 and 2.1.325. The commands and the lines they print were checked on 2.1.325, in a separate config folder. I did not find an official docs page for this plugin.
  • The example is real Claude Code output from a demo folder with made-up data, on 14 November 2026. The plugin was turned on for that run only.
  • The @ClaudeDevs announcement of 2 October 2026, as summarized by dsebastien.net and claudelog.com.

What did it decide on its own for you?

Send me a DM on Instagram with a screenshot of the first note you got, and what Claude Code decided there without asking.

@danielthegoldman
Daniel Goldman

AI-native products, workshops, and automations. Built from everywhere.

© 2026 Daniel Goldman