Why Your Claude Code Bill Is So High (And How to Fix It)
Duncan Rogoff breaks down the four ordinary habits, compacting too early, letting the cache expire, a bloated CLAUDE.md, and untouched defaults, that quietly force Claude Code to re-read your whole chat at full price.
Posted
yesterday
Duration
Format
Tutorial
educational
Views
929
27 likes
57 · 43
Big Idea
The argument in one line.
Claude Code already fights to keep your bill low with prompt caching and lazy-loaded context, and nearly every unexpected cost spike traces back to one of four habits that force it to re-read the whole chat at full price.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're on a Claude Code subscription or API key and your bill or token limit disappears faster than the work seems to justify.
You reach for /compact as a reflex whenever a chat feels slow, without knowing what it actually does to the cost.
You run long sessions with a big CLAUDE.md, several connectors or skills loaded, or frequent switches between Opus and Sonnet.
SKIP IF…
You already understand prompt caching TTLs and token pricing in depth; this stays at a plain-English level, not a pricing deep dive.
You don't use Claude Code; the mechanics here are specific to its caching and session behavior, not general LLM cost advice.
TL;DR
The full version, fast.
Claude Code already minimizes reading cost through prompt caching, lazy-loaded connectors, truncated tool results, and subagents, so most surprise cost spikes come from the user undoing those savings. Four ordinary habits force a full, full-price re-read of the chat: treating /compact as a cost-cutter when it actually costs one full read to produce a summary, letting the prompt cache expire during a break or a mid-task model switch, carrying a bloated CLAUDE.md and too many loaded skills or connectors into every message, and leaving Opus and extended thinking on by default for jobs that don't need them. The fix: stay on one model per job, keep CLAUDE.md under 200 lines as an index, clear or hand off a chat instead of compacting out of habit, and check /usage weekly to confirm it worked.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Defines tokens as Claude Code's billing unit and reframes 'save money' as 'save tokens,' setting up the four traps to come.
00:53 – 01:20
02 · The one habit that matters most: start fresh
The single highest-leverage fix: start a new chat per new job type with /clear, since clearing costs nothing and old chats can be recovered with /rename.
01:20 – 03:16
03 · Why a long chat gets expensive, and how the cache fights back
Claude has no memory between messages, so every reply re-reads the full thread (the 'context'); the prompt cache is what keeps that re-reading cheap, with two real limits: model switches and a TTL (1hr subscription, 5min pay-per-use).
03:16 – 04:34
04 · Three more ways Claude Code already saves you tokens
Connectors load lazily, long tool results get truncated to a reopenable file, and subagents handle messy work on an isolated thread and return only a short summary.
04:34 – 07:12
05 · Trap 1: /compact is not a cost-cutter
/compact reads the whole chat once to produce a summary, so it costs one full read; it's for making room, not saving money, and a handoff doc is often the better third option.
07:12 – 09:14
06 · Trap 2: letting the cache expire
A lunch break is enough to expire the cache, forcing a full-price re-read on return; recent Claude Code versions auto-compact an idle chat right before that happens, and switching models mid-chat breaks the cache the same way.
09:14 – 12:00
07 · Trap 3: what's already loaded before you type a word
A bloated CLAUDE.md rereads in full on every cache miss; treat it as an index, keep it under 200 lines, and lean on skills, scripts, and hooks to keep tool output out of the chat.
12:00 – 13:51
08 · Trap 4: the defaults nobody changed
Opus plus extended thinking is the paid-plan default even for simple jobs, and thinking output costs five times more than input; pick Sonnet, set /effort low, use plan mode, and be specific.
13:51 – 14:32
09 · Checking whether it worked
/usage is the one number that proves the fixes landed: check it this week, apply the four fixes, and check it again next week.
Atomic Insights
Lines worth screenshotting.
Claude Code has no memory between messages, so every reply re-sends and re-reads the entire chat thread from the first message to the newest one.
Prompt caching reads a chat's unchanged portion at roughly a tenth of normal price, which is what keeps a long Claude Code session affordable at all.
The prompt cache expires after about an hour on a monthly plan and after five minutes on pay-per-use or bought credits, and the next message after expiry rereads the whole chat at full price.
/compact doesn't cut your bill directly, it costs one full read of the entire chat to write the summary that replaces it, and only pays off if the next messages stay on the same job.
Switching models mid-chat, including a 'cost-saving' move from Opus to Sonnet, forces a full-price re-read because the prompt cache is tied to one specific model.
A 600-line CLAUDE.md gets sent with every single message, and every time the cache expires Claude pays full price to read all 600 lines again.
Anthropic's own engineers reportedly treat a low cache-hit-rate alert as an incident on the same severity level as a Claude Code outage.
Connectors load lazily: only the connector's name sits in context at first, and its full instruction set loads only once Claude actually uses it.
A skill works the same way as a connector: only a one-line description loads until the job actually needs the full instructions inside it.
A skill that wraps a script lets a computation run locally instead of Claude reading every row and reasoning through the math, which is dramatically cheaper than redoing the work from scratch each time.
Extended thinking output costs five times more than input tokens, so leaving thinking on for a simple job pays premium rates for reasoning the task never needed.
A vague prompt like 'fix the bug' makes Claude scan files to find it, and every file it reads gets added to the chat's running cost, while naming the exact file skips that scan entirely.
/usage breaks down exactly what each chat, subagent, and connector used, making it the only real way to confirm a cost-saving habit actually worked.
Takeaway
Four habits that quietly cost you tokens.
TOKEN COST CONTROL
Every Claude Code bill spike traces back to forcing a full re-read of the chat, so protecting the prompt cache across model choice, session length, and file size is the entire game.
01What tokens are and what costs you money
A token is Claude Code's billing unit, roughly the size of a short word, and it draws down either your plan's token limit or your API bill directly.
Framing the problem as four things quietly costing you money means the fix isn't a new tool, it's breaking four small habits.
02The one habit that matters most: start fresh
/clear wipes a chat back to zero at no cost, and /rename lets you recover an old chat from the sidebar later if you need it.
The rule of thumb worth remembering above all others: a fresh chat per new job type, not per message.
03Why a long chat gets expensive, and how the cache fights back
Claude remembers nothing between messages, so every reply re-reads the entire chat from the first message onward, which is why longer chats cost more.
Prompt caching reads the unchanged part of a chat at roughly a tenth of normal price, which is what makes long sessions affordable at all.
The cache breaks on a model switch and expires on a timer, an hour on a subscription plan, five minutes on pay-per-use or bought credits.
04Three more ways Claude Code already saves you tokens
Connectors load lazily, with only the connector's name in context until Claude actually uses it, not its full instruction list.
A giant tool result gets truncated in the chat and saved to a separate file Claude can reopen only if it needs to.
Subagents handle messy, multi-step work on an isolated thread and return just a short summary, but they still cost tokens, so they're worth it for big jobs only.
05Trap 1: /compact is not a cost-cutter
/compact reads the entire chat once to write a summary that replaces it, so it costs one full read, not a cheaper message.
Use /compact only to make room, when a chat is nearly full mid-task or moving to the next phase of a long project, not to save money.
A handoff doc, a Claude-written file covering the goal, what's done, what's left, and what didn't work, costs about the same as /compact but lets you read and edit what survives before starting fresh.
06Trap 2: letting the cache expire
Stepping away for a break is enough: the cache expires on its timer, and the next message back rereads the whole chat at full price.
Recent Claude Code versions auto-compact an idle chat right before the cache would expire, so the comeback message reads a short summary instead of the full thread.
Switching models mid-chat, even a 'downgrade' from Opus to Sonnet meant to save money, breaks the cache because it's tied to one specific model, so the next message pays full price anyway.
07Trap 3: what's already loaded before you type a word
CLAUDE.md loads into every single message automatically, so a 600-line file gets reread at full price every time the cache expires; keep it under 200 lines.
Treat CLAUDE.md as an index, a trunk holding only top-level rules, with each branch a separate file Claude opens only when a job needs it.
A skill that wraps a script lets a computation run locally instead of Claude re-deriving it from scratch every time, which is dramatically cheaper once built.
A hook can trim a wall of passing test output down to just the failures before it ever lands in the chat and gets reread on every later message.
08Trap 4: the defaults nobody changed
A paid plan defaults to Opus with extended thinking on, and thinking output costs five times more than input tokens, so a simple job pays premium rates for reasoning it didn't need.
Pick Sonnet for routine builds, set /effort to low for easy jobs like renaming a file, and use plan mode to agree on an approach before Claude touches anything.
A vague prompt like 'fix the bug' makes Claude scan files to find it, adding every file it reads to the running cost; naming the exact file skips the scan entirely.
09Checking whether it worked
/usage is the one number that confirms whether the four fixes actually worked, showing exactly what each chat, subagent, and connector used.
Check /usage this week as a baseline, apply the four fixes, then check it again next week: a lower number means more room on a plan or a lower bill on an API key.
Glossary
Terms worth knowing.
Token
The billing unit Claude Code charges for, roughly the size of a short word; it draws down a monthly plan's limit or, on an API key, converts directly to dollars.
Context
Everything Claude re-reads on a given message, which is the entire chat thread from the first message to the newest, since Claude itself remembers nothing between messages.
Prompt cache
Claude Code's short-term memory for a chat: it stores what was already read so the next message only has to read what changed, at roughly a tenth of the normal per-token price.
/compact
A Claude Code command that reads the whole chat once, writes a short summary, and replaces the chat with that summary to free up room, not to cut your bill.
/clear
The command that wipes a chat back to zero at no cost, the standard way to start a new job without carrying forward tokens from the last one.
/rewind
A command that rolls a chat back to before a wrong turn, instead of adding more messages trying to talk Claude out of a mistake.
Handoff doc
A file Claude writes summarizing a job's goal, progress, decisions, and dead ends, meant to be read into a fresh cleared chat instead of carrying the old thread forward.
Connector
How Claude Code plugs into an outside app like Slack or GitHub; only the connector's name loads at first, with its full instruction set pulled in only when it's actually used.
Subagent
A helper chat Claude Code can spin up to do a messy, multi-step job on its own thread, returning only a short summary to the main chat instead of the whole working process.
Skill
A step-by-step instruction file Claude Code only opens when a job needs it; until then, only a one-line description sits loaded in context.
Hook
A filter that trims a long tool result, like hundreds of lines of passing test output, down to just what matters before it ever lands in the chat.
/effort
A Claude Code setting for how hard Claude thinks on a given task, with lower settings suited to simple jobs like renaming a file.
Plan mode
A Claude Code mode where Claude proposes a plan and gets agreement before touching any files, often paired with a bigger model for planning and a cheaper one for the build.
/usage
The Claude Code command that shows exactly what each chat, subagent, and connector used, the only real way to confirm a cost-saving habit is actually working.
03:05linkAnthropic's engineering write-up on building Claude Code
Quotables
Lines you could clip.
01:25
“Claude is like a brilliant contractor with total amnesia.”
sharpest one-line definition of context in the whole video, instantly explains why chats cost more over time→ TikTok hook↗ Tweet quote
02:10
“A cache is just short-term memory.”
simplest possible definition of a complex billing mechanic→ newsletter pull-quote↗ Tweet quote
05:13
“It's like asking your assistant for meeting minutes. Somebody has to read the whole transcript to write them.”
reframes the most misused command in Claude Code in one analogy→ IG reel cold open↗ Tweet quote
12:08
“Running Opus on every job is like sending your best surgeon to take your temperature.”
vivid and funny, nails the default-settings cost mistake in one line→ TikTok hook↗ Tweet quote
14:31
“The cheapest message is the one Claude already remembers.”
closing mantra line, works standalone as a pull quote→ newsletter pull-quote↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogy
Everybody says you need GitHub repos, extra add -ons, or a long list of hacks to save tokens, but really, you don't. Anthropic already built Cloud Code to save you tokens, and Cloud Code gets better at cutting your costs with each new release. But there are four things that are quietly costing you money, and you're probably doing them right.
Now I'll show you all four and the best ways to get the most out of your cloud plan or your API usage. And at the end, I'll show you the one number that tells you if any of it is actually working first one word, because it matters for this whole video. That word is tokens.
A token is the unit quad charges you for, and it's about the size of a short word on a monthly plan. Tokens use up your limit. If you pay per use with an API key tokens are dollars.
So when I say cost, I mean. tokens. So here's what actually saves you money.
If you understand one concept from this video, make it this because it's the base for everything else. The longer your chat runs, the more every new message costs. So the fix is to start a fresh chat for each new job type slash clear and quad code wipes your chat so you start at zero.
Clearing costs you nothing. If you want your old chat back later, just rename it by typing slash rename. Then you can find it in the sidebar and open it up again.
So here's how Claude code actually works and why a fresh chat saves you money. It comes down to one fact about Claude. Claude is like a brilliant contractor with total amnesia.
It doesn't remember anything between messages. So every time you send a message, you're really sending the whole thread from your first message your newest one.
Claude reads the whole thread, answers, and forgets it all again. Everything Claude reads on each message is called the context, and that reading is what you pay for. So a longer thread means more reading and more tokens on every message.
That's why your chat gets more expensive the longer it runs. And a file Claude reads early stays in the thread, so you pay for it again on every message after. That's why you clear your chat or make a new one when you start a new job.
All that reading sounds expensive, but Cloud Code has a trick to keep that reading cheap, and it's called the cache. A cache is just short -term memory. Most of your chat hasn't changed since your last message.
So the cache remembers what Cloud Code already read, and it only has to read your new message. Reading from the cache costs a tenth of the normal price or less. But here's the catch.
The cache only remembers your chat as long as nothing changes. Change something like switching to a different model and cloud code has to read your whole chat from scratch again at full price. And the cache doesn't last forever.
On a monthly plan, it forgets after about an hour. If you pay per use or you're using extra credits you bought, it forgets after five. minutes.
Anthropic cares about this a lot. In their own write -up on building cloud code, they say their engineers get an alert when the cache isn't being used enough, and they treat that alert as an emergency the same way they would if cloud code went down. down.
But the cache isn't the only way Cloud Code keeps your reading light. Here are the other tricks, one at a time. The first trick is connectors.
A connector is how Cloud plugs into an app like Slack or GitHub. Each connector comes with a long list of instructions for how to use it. Cloud Code doesn't load all of that at the start.
It loads just the connector's name, and it pulls in the rest only when Cloud actually uses it. So Click the plus, then connectors, and browse for existing connectors to your most frequently used platforms.
There's a lot of them. The second trick is how Cloud handles really long results. Sometimes, Cloud gets back a giant wall of text, like when it checks its own work.
Cloud Code cuts off the long ones and saves the full thing in a separate file that Cloud can open if it needs it. And the third trick is helper chats called subagents. For a big, messy job, Cloud Code can start a helper that does the messy work on its own.
Only a short summary comes back to your chat. But helpers cost tokens too, so they're only worth it for big jobs. So Cloud Code is already working to keep your reading cheap.
Your job is to stop undoing it. And the first thing that undoes it is a command you've probably typed thinking you were saving money. So trap one is treating compact like a diet.
So picture one chat. You're building a landing page and 90 minutes in, you're at message 60 and it's slow. So you try a command called slash compact.
But here's what slash compact is supposed to do. It takes your whole chat, writes a short summary of it and replaces the chat with that summary. Everybody says that shrinks the chat.
it must shrink the bill. But does slash compact actually save you money? So here's what happens.
To write the summary, Cloud has to read the entire chat one more time. And that read is an extra request. Then the summary replaces your chat.
It's like asking your assistant for meeting minutes. Somebody has to read the whole transcript to write them. The minutes are short, but the reading already happened.
If the cache is fresh, that extra read is cheap. But if the cache expired, you pay full price for it, and the summary starts a new cache from scratch. So what is slash compact actually good for?
Making room. There are two times it makes sense. One is when your chat is almost full and you're in the middle of a job you can't restart.
Cloud can only hold so much at once, so something has to get cut down. The other is when you're deep into a long project and about to move on to the next part of it. You can actually tell slash compact what to keep, like slash compact, keep the plan and the file names.
If your next job has nothing to do with this one, slash clear is the better move. Either way, saving money isn't what compact is for. You pay for one big read right now, and the next messages are cheap only if you keep going on the same job.
It can also lose small instructions you gave early on, so check that Claude still remembers the important ones. Claude code also compacts on its own when your chat gets close to full, so you don't need to babysit it. So if you're done with the job, use slash clear.
And if Claude takes a wrong turn, don't type that didn't work, try this, type slash rewind and Claude code takes the chat back to before the wrong turn. And when your job isn't done, but your chat is heavy, there's a third option and it's my favorite, a handoff doc. Ask Claude to write a thorough file that covers your goal, what's done, what's left, the decisions you made and what didn't work.
Then run slash clear and tell your new chat to read that file. That costs about as much as Compact or a little more because Claude has to read your chat and write the doc. But you get to read the doc and fix it first, you decide exactly what survives, and you can keep it for next week.
But there's a problem. There's actually a clock running behind every chat that can wipe out all of this, and you've probably never seen it. Trap number two is the cash expiring.
So that clock is the cash expiring. Let's go back to our landing page example. You've been working on your landing page all morning and you're not finished.
So you go to lunch. When you come back, you ask one tiny question, but what you don't realize is that this message costs more than the last one. Why?
Because the cache expired. So when it expires, your next message rereads your entire chat at full price. And after that, the cache is fresh again.
Anthropic added a fix for this. Recent versions of Cloud Code compact a big chat that's been sitting idle for close to an hour right before the cache expires. So when you come back, Cloud Code reads the short summary and not the whole chat.
That's a good example of Cloud Code getting better at cutting your costs with each release. But it's still compact, so it still has to read the whole chat once, and it can still lose small details. Another thing that wipes the cache is switching models.
So Cloud comes in different models like Opus and Sonnet. Opus is the bigger and more expensive one, and Sonnet handles most jobs for less. The cache is tied to one model, so say you're an hour into a chat on Opus, and you switch to Sonnet to save money.
Sonnet can't use Opus's cache, so it reads your whole chat from scratch. at full price. That's a surprise because switching to a cheaper model sounds like saving, so pick your model before the job starts and stick with it.
There's also a setting that switches models for you. If you type slash model opus plan, all one word, While you're in plan mode, Claude uses Opus, the bigger model, to make the plan.
When it's time to build, Claude switches to Sonnet, the cheaper model, to do the building. It's useful, but every switch costs that same one -time read. So if the job's done, when you walk away, clear it.
But if the job's not done, a big chat may get compacted or the first message back costs more. The third of the four things quietly costing you money happens before you type a single word. Let me show you what I mean.
Trap three is what's already loaded. So you open a fresh chat and you type, hi. But what you don't realize is that a big file just got loaded into your context automatically.
That file is Claude .md, which is the file Claude reads at the start of every chat. Say yours is 600 lines of rules. Claude code sends ClaudeMD with every message.
So while the cache is fresh, it's cheap. But every time the cache expires, Claude pays full price to read all 600 lines again. So anthropic stocks say to keep Claude .md under 200 lines.
The way to do that is to use your Claude .md file like an index. Think of it like a tree. Claude .md is the trunk and it holds only your highest level rules.
Each branch is a separate file for one kind of job and Claude .md tells Claude which branch to climb. Or it can point to one index file that lists where everything else lives. Either way, Cloud only opens a branch when the job needs it.
Skills work the same way. A skill is a text file of step -by -step instructions that Cloud only opens when it needs them. Until then, just a one -line description is loaded.
And skills have a much... bigger trick. A skill can also carry a script, which is a small program that does one job.
So say every Monday you pull last week's numbers out of a spreadsheet. Instead of Claude reading every row and working out the totals from scratch, a script adds them up on your computer. Never reads the code, it only sees the result like 42 sales up from 38.
That's way more efficient than having Claude write new code and new rules every time. So once you have a process that works, tell Claude, Turn this into a skill.
You create it once and next time you use it for almost free. The same thing happens with connectors and skills. It's easy to end up with too many of both and each one adds a little to every session.
So type slash MCP and turn off the connectors you aren't using. Then type slash context, which shows you what's taking up space in your chat. And every so often run the slash doctor command, which is a checkup for your setup and it flags problems you can clean up.
Next up is the text Claude gets back when it checks its own work. When Claude builds something, it runs tests, which are quick checks that your project still works. A test can print hundreds of lines and most of them just say, passed.
Cloud Code only cuts results that are much longer than that. So every one of those lines goes into your chat and gets reread on every message after. The fix is something called a hook, which is a little filter that trims those results before Cloud sees them.
You don't have to build it yourself. Just tell Cloud, write me a hook that trims test results down to just the failures. The last of the four things quietly costing you money is a default.
A default is a setting Cloud Code chose for you before you. you ever opened it. And two of them are quietly making every chat cost more.
So trap four is the default. On a paid plan, cloud code starts on Opus and thinking is on. Thinking is cloud writing out its reasoning before it answers.
Everything cloud reads counts as input and everything cloud writes counts as output. That reasoning counts as output and output costs five times more than what input does. So running Opus on every job is like sending your best surgeon to take your temperature.
It works, but you're paying for skills the job didn't need. So pick Sonnet for the build. Type slash effort, which is how hard Claude thinks, and pick low for easy jobs like renaming a file.
Switch to plan mode so Claude agrees on a plan with you before it touches anything. And be specific. If you type fix the bug, Claude scans files to find it, and every file it reads gets added to your chat.
But if you type fix the date bug in the checkout file, it goes straight there. The words you type are a tiny part of what Claude reads, so shorter prompts barely matter. What you point Claude at matters far more.
So that's the four things, and here's the whole thing. Every one of these four things comes from the same place, making Claude re -read more than it has to. a model switch, a long call at MD, and a lunch break all add reading.
So the goal is simple. Keep the cash working for you. Stay on one model for the whole job, keep going while the cash is fresh, and clear your chat when the job is done.
If you pay per use, that window is only five minutes. So take your breaks between jobs and not in the middle of one. And skip compact unless you really need the room.
But how do you know any of it is actually working? Well, there's one number that tells you and it's slash usage.
It shows you what each chat, subagent, and connector actually used. So here's what to do this week. Type slash usage today and look at where your tokens went.
Then do the four things. A week from now, type slash usage again. On a monthly plan, a lower number means more room before you limit.
On an API key, a lower number means a lower bill. That's how you know it worked. If you're building with Claude every day, I co -founded the Claude Code Club.
It's $9 a month and the link's in the description. The cheapest message is the one Claude already remembers.
The Hook
The bait, then the rug-pull.
Duncan Rogoff opens by knocking down the usual advice, extra repos, add-ons, a pile of hacks, and argues Claude Code already works to keep a session's reading cheap. The real problem, he says, is four ordinary habits that quietly force it to re-read the whole chat at full price anyway.
Frameworks
Named ideas worth stealing.
04:34list
The Four Hidden Cost Traps
Treating /compact like a diet
Letting the prompt cache expire
What's already loaded before you type (CLAUDE.md, skills, connectors)
Expensive defaults (Opus + thinking left on)
The video's spine: four ordinary Claude Code habits that quietly force a full re-read of the chat at full price.
Steal forany 'hidden costs' explainer structure for a technical tool
03:16list
Three built-in cost savers
Connectors load lazily (name only, full instructions pulled in on first use)
Long tool results get truncated and saved to a file Claude can reopen
Subagents do messy work in an isolated thread and return only a short summary
Before naming the user's mistakes, the video credits Claude Code's own architecture for already keeping reading cheap.
Steal forany cost-optimization explainer that wants to credit the platform before blaming the user
CTA Breakdown
How they asked for the click.
VERBAL ASK
14:31product
“I co-founded the Claude Code Club. It's $9 a month and the link's in the description.”
Single soft mention in the last 15 seconds of a 14.5-minute video, backed by a free companion guide linked in the description rather than pitched on camera.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A professional motion designer hands his entire YouTube edit, script to graphics to quality check, to Claude Code and breaks down what it actually costs.
A reverse engineering tool called REA jumped from 13,000 to 149,000 GitHub stars in five days, and this creator uses it to make the case for what still has value once any app can be copied.
A week-old open-source project called ArtCraft cloned Photoshop, Premiere, After Effects, Illustrator, Lightroom, InDesign, and Acrobat for free, mostly by pointing Claude Opus at the problem, and this breaks down which clones are actually ready to use today.
Every night at 1am, 3am, 4am, and 5am, a set of Claude Code agents build something, grade themselves against a rubric, and write their own rules before the next run.
One creator claims Claude Opus 5.5 replaced the $1,500-a-month editor who used to make his Instagram Reels, and walks through the eight-step system that does it.
00:26
00:37
00:50
00:53
01:06
01:25
01:36
01:46
01:57
02:08
02:19
02:39
02:51
03:02
03:13
03:23
03:45
03:56
04:07
04:17
04:30
04:50
05:04
05:13
05:24
05:36
05:47
06:04
06:09
06:21
06:32
06:43
06:55
07:06
07:28
07:36
07:50
08:01
08:13
08:24
08:40
08:46
08:57
09:08
09:30
09:41
09:52
10:03
10:14
10:25
10:37
10:48
10:59
11:10
11:21
11:32
11:43
11:54
12:16
12:27
12:38
12:49
13:01
13:12
13:22
13:34
13:45
13:55
14:16
09:44
14:01
08:47
Video of the Day15:25
07:20
Video of the Day13:56