Go Install This Right Now: The Quiet Second Claude That Reads Over Your Shoulder
Go Install This Right Now: The Quiet Second Claude That Reads Over Your Shoulder
One built-in switch in Claude Code. Off by default. Here’s what it does that the docs don’t tell you.
I’ve watched Claude make a perfectly reasonable decision, mention it once in the middle of a long answer, and keep going. I didn’t read that sentence. Neither do you.
That’s not a character flaw — that’s what delegating means. But somewhere in step 14 of a long task is the call you would have argued with, and it’s sitting in the scrollback between a test run and a file diff.
There’s a switch in Claude Code that catches those, and it’s off by default. Turn it on. It takes one command:
/plugin enable cc-plugin-you-should-know@builtinThe short version — if this box is all you read
- It’s a second, quiet Claude that rereads your session during long tasks and posts one short line above your prompt when something with real consequences got buried.
- Most of the time it says nothing — by design.
- It has no tools and never interrupts the main agent. It can’t touch your files; it can only talk to you.
- A real one from my sessions: a few tests had been quietly pointed at an old version of a page to get a deploy through. They passed — and no longer tested the page anyone uses.
- It’s built in: nothing to update or maintain. Off is the same command with
disable. - It costs about 11% on top of a long session — roughly 12 cents a check on Opus. Short tasks cost nothing; it never fires on them.
- The catch: it needs telemetry on, and it isn’t offered on Bedrock, Vertex, or under zero-retention/HIPAA terms.
$ tail -f session.log
What you actually see
● Updated 3 files · wrote summary…
One line, about twenty words, tagged one of two ways:
Learn more opens a hundred-word explainer. From there, Chat in main session drafts a prompt about it into your prompt box, so you can hand it straight to the main Claude.
Five out of six notes I took into the session were Heads up. The value isn’t teaching. It’s consequences I’d have missed mid-task.
$ man you-should-know
What the docs don’t tell you
Anthropic documents the switch, not the behavior. I read how the shipped plugin works; here it is in plain words. (Current as I write this — it ships inside Claude Code, so any release can change it.)
On every sixth step Claude takes inside one turn — a step being one tool call or one reply. Six steps into a long task, then twelve, then eighteen. Each look is a single, separate model call that rereads the conversation you already have, then speaks once or stays quiet. A quick back-and-forth never reaches six, so it never fires.
Mostly when to shut up. Default to saying nothing. Skip trivia, skip anything you already discussed or clearly understand, skip anything it isn’t confident about. The bar is a real consequence: money, time, wasted work, a wrong result, or a decision you’re in the middle of.
My favorite line in its instructions: “interesting is not the same as important.”
It won’t repeat a recent note, or a topic you closed with Knew this already. That button is how you stop a topic coming back.
Ignore a few notes in a row and it starts skipping checks. Answer any note and it resets.
The exact numbers
It remembers roughly the last fifty notes it showed you. A note you type past clears after a couple of prompts. After three ignored in a row it skips one check, then two, then four, up to sixteen.
It’s only as good as its ability to stay quiet.
$ grep -ri “heads up” ~/sessions/
Does it actually catch anything?
Yes. These are real notes from my own sessions, generalized so they don’t point at any one project. Every one was about something Claude had already done or said in the scrollback — I just hadn’t taken it in.
The tests that stopped testing. To get a deploy through, a few tests had been pointed at an old version of a page. They passed. They no longer checked the page anyone actually uses.
The fix that only lived in one place. The fix for a live problem was on my desktop and nowhere else. The copy everyone else uses was still broken.
The article I didn’t approve. After I’d signed off on a draft, more review rounds rewrote its opening, its subtitle and several of its stories. The version about to ship wasn’t the one I approved.
“Tested” meant “tested on my machine.” A new button had only been tried in my own browser — never in the page readers would actually load.
One more catch
The docs nobody will find. A new how-to got saved somewhere my own tooling doesn’t index, so future sessions looking for it would come up empty.
Most of these are about things being out of sync — a fix in one place but not another, a stale copy, a step done in the wrong order. That’s what slips in long, late sessions spread across several machines.
That’s exactly how I work, which is why it earns its keep for me. It’s worth less if you read every line Claude writes, or your tasks are short.
The receipts
About 120 notes shown in three and a half weeks. I took 30 of them into the main session — still roughly 1 in 4. My tracking misses notes taken while Claude is mid-turn, so the real number is a little higher.
$ cat FINE_PRINT
The honest part
I’d tell every Claude Code user to turn this on. I’d also tell them what they’re turning on.
- It needs telemetry. If you’ve set
DISABLE_TELEMETRYorCLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, it won’t show up. Claude Code’s telemetry goes to Anthropic and a third-party logging service, per Anthropic’s own data-usage docs. - Its telemetry doesn’t carry the note text.
- It isn’t offered everywhere.
--safe-modedoesn’t stop it. Neither does--bareor turning off all hooks — those affect plugins you installed, not built-in ones. The/pluginswitch is the only off switch.- You can’t tune it. How often it looks, how high its bar is, and which model it uses are all fixed. Your levers are on/off, Knew this already, and ignoring it.
- It’s per machine. The switch lives in your local settings — turn it on everywhere you work.
What its telemetry sends, and where it isn’t offered
Telemetry, from the code: which button you pressed, timings, token counts, the version, and a one-word category. Not the note itself. The check is a normal model call to Anthropic on your plan; the plugin can’t make network calls of its own.
Not offered: on Bedrock, Vertex or Foundry; for organizations with zero-data-retention or HIPAA terms; in Claude Desktop’s agent mode. It’s also still rolling out, so some people won’t see it yet.
$ /plugin enable cc-plugin-you-should-know@builtin
Go turn it on
/plugin enable cc-plugin-you-should-know@builtinDo a real day of work with it — something long — and judge it by one test: when it spoke, was it worth reading? Old news? Press Knew this already. Noise? Ignore it and it backs off.
And if it ever catches the thing you’d have shipped broken, you’ll stop asking whether it was worth one command.
$ cat cost.csv
What it costs
Will it cost you? Maybe. It depends on how you pay for Claude.
Pro or Max subscription
Nothing extra.
The checks count toward your usage limit, like any other Claude Code call. If you’re regularly hitting that limit, this plugin isn’t your problem.
Paying per token
Real money.
The API, or any plan billed on usage. At API list prices, my sessions would have added about $80 a month. Yours scales with how long your sessions run.
Pro tip
Turn it on for new work.
That’s how I run it: on when I’m building something new — where the important calls get buried — and off for routine work.
The choice is yours. Here’s the data.
I measured it across two and a half weeks of my own long sessions: 458 checks, each one’s tokens read from Claude Code’s debug log and priced at Anthropic’s published API rates.
| What | Measured |
|---|---|
| One check, in an Opus 5.5 session | ~$0.12 (most expensive: $0.26) |
| One check, in a Sonnet 5.5 session | ~$0.06 |
| Checks that said anything | 1 in 7 |
| Added to a typical long session | ~$0.85 (my most expensive: $6) |
| Added to the session’s own cost | +11% (most sessions: 8–18%) |
| My total, two and a half weeks | $48 — about $80 a month |
You pay for the silence. Six of every seven checks say nothing, and a quiet check costs nearly as much as one that speaks — about a dime either way. That’s the price of a reviewer with a high bar: it has to read everything to decide most of it isn’t worth your time.
The good news runs the other way. It only looks during long turns, so it only spends money in the sessions where things get buried. Quick back-and-forth work costs nothing, because it never fires. And in a Sonnet session each check costs half as much.
Where a check’s money goes
Each check rereads the conversation from the prompt cache — on my sessions, usually 110,000 to 370,000 tokens. That’s the cheap part: about a third of the price. Half the price is about 16,000 tokens that aren’t cached and get sent fresh every time. What it writes back is under a tenth. Longer conversations mean bigger rereads, so checks get pricier as a session grows.
Measure your own — and the trap I fell into
It isn’t in your transcript. The check runs outside the session log, so tools that estimate spend from transcripts can’t see it. Start Claude Code with --debug-file <path> and the debug log records each check’s token counts.
Check your rate table. My own dashboard first said $84. It was pricing the new models at the previous generation’s rates; the real number was $48. And debug logs hold file paths and hook output — delete them when you’re done.
So, is it worth it?
I haven’t formally rated its notes yet, so this isn’t a hit-rate score. Here’s what I can measure: in the sessions where I tracked cost, I handed 19 of its notes to the main Claude to act on — about 1 in 3 — at roughly $2.50 per note. (My tracking misses notes taken mid-turn, so the real cost per note is a little lower.)
$ echo $VERDICT
The Verdict
Any one of the catches above would have cost me more than $2.50 to find the hard way.
It’s staying on — for the work that’s new.