DEV Community: Станислав Кириченко

86 Billion Tokens, One Click, and No One to Talk To: How I Lost My Anthropic Account

Станислав Кириченко — Thu, 12 Nov 2026 12:04:30 +0000

A few months ago I published an article here about how Claude Code changed the way I work — how 20 years of SEO led me to prompt engineering, MCP servers, and building open-source tooling on top of Claude. It was an optimistic piece. This one isn't.

Two weeks ago I opened Claude and saw a restricted page instead of my workspace. My account — one personal organization, 8 months of daily work, roughly 86 billion tokens of research, client projects, and tooling — was suspended for "unusual activity."

I want to be clear about what this post is and isn't. It isn't a rant, and it isn't a demand. It's a description of a process that, in my experience, gives a paying user no realistic way to be heard. If someone from Anthropic reads it — even better.

What I lost

Context matters here, so briefly: I've been in technical SEO for 20 years. For the last 8 months, Claude wasn't a toy for me — it was my research environment. I built and open-sourced global-design-skill, an MCP-based design rule system. I ran large-scale SEO research through Claude Code. I prototyped tools I'd been imagining for years and never had the engineering bandwidth to build.

86 billion tokens sounds like abuse. I get that. But it was real, visible, consistent work — the same kind of usage, day after day, month after month. If my account history were a resume, it would be the most boring resume in the world.

What those 86 billion tokens actually produced

Here's the part I wish someone at Anthropic would look at before calling my activity "unusual." In these same 8 months I published a series of open research preprints on Zenodo — every one with data, code, and replication scripts attached. All of them were built with Claude as my main research environment:

A study of nine popular prompting techniques (chain-of-thought, chain of draft, and friends) tested against a plain instruction across three real LLM stacks — including Claude Haiku through Claude Code. Result: on deployed stacks, fancy prompting mostly loses to a clear instruction, and the metric often rewards degenerate answers. 7,937 valid model responses, all published.
Two blind-scored studies on whether covering a search query's intent predicts ranking in Yandex results — the second one preregistered before data collection. Spoiler: it doesn't. Median correlation with rank was statistically indistinguishable from zero.
A snapshot study of what separates Google's top 20 from positions 21–50 — 7,370 URLs, 160 queries, with an LLM assessor scoring intent coverage blind to rank, and a comparison against Yandex.
A measurement of how stable AI search citations actually are — whether Perplexity cites the same sources when you ask again, or ask the same thing in different words. 200 answers, bootstrap over all resamples.
A comparison of AI citation factors across Google AI Overviews, AI Mode, Perplexity, and Yandex Alice — 2,400 Russian-language queries, 38 ranking factors from vendor SEO studies tested against what actually gets cited.
Two studies on machine-text detection moved to Russian and to the SEO genre — what perplexity-based detectors actually measure, and how formatting instructions alone shift machine-likeness scores when the text's origin stays fixed.

I'm not listing these to brag. I'm listing them because this is what "unusual activity" looked like on my account: a researcher using the tool heavily, in the open, and publishing everything — including the hypotheses that failed. The token volume that probably triggered the flag is the research. If my usage pattern looked anomalous, that's what the anomaly was made of.

The ban, and the page that says nothing

The restricted page told me my organization was suspended for unusual activity and that I could request a review. I did. I wrote honestly: I said I couldn't rule out that the trigger was my login geography. My work involves accessing client server environments in various regions — that's a normal part of SEO and site maintenance. If a session routed through client infrastructure looked like access from an unsupported region, that's on my setup, not on any intent to bypass restrictions. I never used Claude for anything against the usage policy. It was all my own work.

Then I clicked through to the appeal form — and found a dead end. The form showed my organization's status: suspended, "final decision." No active buttons. No text field. No way to submit anything at all.

Since then I've sent more than 6 emails. Every single one came back with the same automated reply: a link to the appeals form. The form that has no way to appeal.

That's the loop. The email bot sends you to the form. The form says the decision is final. There's no human anywhere in the circle.

What actually bothers me

If Anthropic's Trust & Safety team looked at my account and decided the geography pattern or the volume was a genuine policy violation — I can disagree, but I can accept it. Companies get to enforce their rules.

What I can't accept is the architecture of the process:

You never learn what you did. "Unusual activity" is not a reason. It's a category. I can only guess — was it the logins? The volume? Something else entirely? A user who doesn't know the actual reason can't fix it, can't explain it, and can't even properly apologize for it.
The appeal channel is a façade. An appeals form that displays "final decision" with no input fields isn't an appeals process. It's a status page. And an email address that only replies with a link to that status page isn't support — it's a loop.
Eight months of history counts for nothing. No human glanced at my usage pattern and thought, "hold on, this looks like a heavy but legitimate user who maybe logged in from a weird place." Or if one did, there's no trace of it — no question asked, no clarification requested, nothing.

I'm not entitled to an account. But I'd argue any paying user is entitled to one human conversation before the door closes.

The lesson nobody wants to learn

Here's the uncomfortable part, and it applies to you even if you never get banned: if your work lives on a platform, the platform's appeals process is part of your infrastructure. I had 8 months of research context, project history, and workflows tied to a single account with a single provider. I exported nothing. I assumed the door would stay open because I was a good customer.

Good customer, bad customer — the automation doesn't care, and right now there is no reliable path past the automation.

So: export your context. Mirror your prompts and tooling. Keep your workflows portable across models. Not because Anthropic is evil — they aren't, and Claude is genuinely the best tool I've used — but because "the best tool" and "a tool you can lose silently in one click" are the same tool.

What I'm asking for

If anyone from Anthropic reads this, my case is TS-01a19256-f63b-7354-9328-5982c459399b. I'm not asking for special treatment. I'm asking for a human review where I can actually explain my situation once — and if the answer is still no, at least I'll know a person said it.

And if you work on Trust & Safety tooling anywhere: please build an appeals flow that includes a text field. That's it. A text field, and a human on the other end, at least once per case.

I still think Claude is a remarkable product. I spent 8 months building on it and I'd go back tomorrow if the door opened. That's exactly why this process stings — you don't write posts like this about products you don't care about.

If you've been through a similar suspension — with Anthropic or any other AI provider — I'd genuinely like to hear how it ended. Comments are open.

From Shell Scripts to MCP Servers: How SEO Broke My Brain (in a Good Way)

Станислав Кириченко — Tue, 26 May 2026 17:56:05 +0000

I've been doing technical SEO for 20 years. My first "automation" was a shell script
that pinged Google every hour to check if my site got indexed. We've come a long way.

For most of those two decades, automation meant one thing: Python. Scrapers,
crawlers, log analyzers, bulk content pipelines. I got comfortable with the idea
that if something was repetitive, you wrote a script for it. Ship it, forget it,
move on.

Then two years ago something broke that model completely.

The thing about prompt engineering nobody tells you

When LLMs got good enough to be useful for real work, my first instinct was to
use them as fancy autocomplete. Better keyword research. Faster content briefs.
Smarter meta description generation.

That phase lasted about three months.

The shift happened when I stopped asking "what can I use AI for" and started
asking "what does a good AI instruction system look like." That question is
completely different — and it pulled me back into something that felt a lot
like software engineering.

Except instead of writing functions, I was writing behavior specifications.

# Bad prompt (autocomplete thinking)
"Write me 10 title tags for a page about industrial pumps"

# Good prompt (behavior specification thinking)  
"You are auditing title tags for B2B industrial equipment pages.
Flag any title that: uses generic verbs (Buy, Get, Find), 
exceeds 60 chars, or leads with brand name instead of product category.
For each flagged item, explain the specific SEO risk and rewrite it."

The difference isn't just style. The second one is reliable. You can run it
on 500 pages and get consistent, actionable output. The first gives you
something different every time.

Vibe coding changed how I prototype

Here's where it gets weird. Around the same time I was getting serious about
prompt engineering, I started using Claude Code and Cursor for my actual
Python work.

I'm not a great developer. I'm a good enough developer — the kind who can
read code, understand what it does, fix obvious things, but struggles to
architect something from scratch without spending too long on decisions that
don't matter yet.

Vibe coding fixed that for me. Not because AI writes perfect code (it doesn't),
but because it eliminates the blank page problem.

My workflow now:

Describe what I want to build in plain language
Get a working skeleton in minutes
Actually understand what was built before touching it
Iterate from something real instead of from nothing

The understanding step is non-negotiable. I've seen people vibe-code themselves
into a codebase they can't maintain because they never stopped to read what was
generated. That's not automation — that's technical debt with extra steps.

When the two things collided

Six months ago I was trying to solve a specific problem: every time I used an
AI agent to help with UI/UX work (landing pages, SaaS app layouts, design audits),
it would generate the same bland outputs. Centered hero. Purple-to-indigo gradient.
Generic cards grid. The AI "safe choices" problem.

My SEO brain said: this is a rules problem, not a capability problem.
The model knows good design — it just doesn't have a consistent framework
to apply it in my context.

So I built one. A set of design rules, blueprints, and industry-specific
guidelines that I inject into every design-related task. OKLCH colors.
Proper motion/react patterns. Sector-specific banned patterns
(different rules for a B2B SaaS vs a health platform vs an e-commerce site).

Then I made it a proper MCP server so any AI tool could use it:

# Any agent can now call:
get_sector_context("b2b-products")
# → Returns required trust signals, banned patterns, conversion elements
# specific to B2B industrial/enterprise design

check_banned_patterns("saas", "purple gradient hero with centered H1")  
# → Returns violations with specific rule citations

The result is global-design-skill
— an open-source design OS for Claude Code, Cursor, Copilot, and Windsurf.
It detects your business sector automatically, applies the right rules,
and the longer you use it the more it calibrates to your feedback.
All local, no telemetry.

What 20 years of SEO actually taught me about AI systems

Looking back, SEO was always about constrained optimization — you can't
control the algorithm, so you control everything you can. Structure.
Signals. Context. Consistency.

Prompt engineering is the same discipline applied to a different system.
You can't control what the model knows, but you can control:

The frame — what role it's playing, what rules apply
The context — what it knows about your specific situation
The output format — what a good answer looks like structurally
The failure mode — what it should say when it doesn't know

The people who are best at this, in my experience, aren't necessarily
great coders or great writers. They're people who are good at thinking
in systems — seeing the inputs, the process, and the expected outputs
as something you can specify and iterate on.

That's a skill SEO built in me without me realizing it.

What's next

I'm currently learning a lot about:

MCP server design — what makes a good tool vs a bad one for AI agents
Context window economics — how to pack maximum useful signal into minimum tokens
Sector-specific AI behavior — same task, different industries, completely different right answers

If you're working at the intersection of AI tooling, automation, and domain
expertise — I'd love to connect. This space is still early enough that most
of the interesting problems haven't been solved yet.

What's your workflow when you're using AI tools for work that's specific to
your domain? Do you inject domain context explicitly or let the model figure it out?