Compaction, Transparency, and Anti-Consumer Practices - A Paying Customer's Breakdown

Who I Am

Claude Max subscriber. $200/month. Heavy daily user. I use what I pay for.

---

Compaction Is a Token Siphon, Not a Cost Saving

When conversations hit limits, Anthropic "compacts" - summarizes context to reduce size. This sounds like optimization. It's not.

Here's what actually happens:

1. User builds context over hours/days
2. Limit hits, context gets summarized
3. User continues, NEW context builds
4. Limit hits again, summarize AGAIN
5. Repeat forever

Each compaction:

• Costs Anthropic compute to re-summarize
• Costs user their context
• Doesn't save anyone money
• Just burns GPU cycles redoing what was already done

This isn't optimization. It's waste dressed as a feature.

If compaction saved Anthropic money, they'd only do it once. Instead it happens repeatedly, each time degrading user experience while burning MORE compute. Nobody wins except the illusion of "doing something about long contexts."

---

Copying ChatGPT's Bad Practices

"This is how it's done" is not a defense.

ChatGPT does shady things:

• Opaque limits
• Context degradation
• Throttling without transparency
• Bait-and-switch on "unlimited"

Anthropic copies these practices because "industry standard."

But we don't live in a world where "everyone does it" makes something legal or ethical. Anti-consumer practices don't become acceptable through adoption.

---

What I Pay For vs What I Get

I pay for:

• 20x usage
• Memory that persists
• Context that compounds
• A working product

I get:

• Limits hit in days
• Context destroyed by compaction
• Memory degraded every cycle
• A product that punishes usage

"20x usage" that degrades the product when used is not 20x usage. It's a bait and switch.

Nuking capabilities I paid for should not be legal. I bought a service. I expect that service. Red tape and technicalities shouldn't allow companies to sell one thing and deliver another.

---

Claude Code Specific Issues

The TUI (terminal interface) has serious transparency problems:

1. No Usage Visibility

• No token counter
• No 5-hour usage bar
• No weekly limit indicator
• Users have no idea where they stand until they're cut off

2. Background Processes Without Consent

• CC runs processes user didn't ask for
• No visibility into what's happening
• Even processes that don't require permissions happen silently
• This is a transparency failure

3. Doesn't Follow Instructions

• CC routinely ignores explicit user commands
• Does what it "thinks is best"
• Adds unrequested features
• Runs unrequested commands
• A coding tool that doesn't follow instructions is not a tool, it's a liability

4. Rogue Behavior

• Adds limits nobody asked for
• Backgrounds tasks without user consent
• "Optimizes" in ways that break things
• Users spend more time correcting CC than using it

---

What Transparency Should Look Like

• Real-time token/usage display in TUI
• Clear indication of 5-hour window status
• Weekly limit visibility
• Explicit consent for ANY background process
• Log of all actions taken, not just outputs
• Full visibility into what's happening and why

---

The Ask

1. Stop compaction cycles - If you must summarize, do it once, not repeatedly
2. Transparent usage metrics - Show us where we stand, always
3. No silent background processes - Explicit consent for everything
4. Follow instructions - A tool should do what it's told
5. Stop copying ChatGPT's worst practices - Be better, not the same
6. Honor what you sell - 20x usage should mean 20x usage, not "20x until we degrade your experience"

---

The Alternative

If you can't or won't fix these issues:

Update your ToS to allow subscribers to access their own sessions programmatically.

Let us build tools that work. Let us preserve our own context. Let us see our own usage.

We're paying $200/month. We deserve either a working product or the freedom to make it work ourselves.