This week Anthropic cut off OpenClaw users from their Claude subscriptions. If you were running an OpenClaw agent on a $200/month Max plan, you were getting $1,000 to $5,000 per day in equivalent API compute. That subsidy is gone.
A post about this is making the rounds, and it introduces a framework I think is genuinely useful — four tiers of AI users:
Tier 1: Spectators. Still watching. Reading about AI, not using it.
Tier 2: Casual users. Rewriting emails, summarizing docs. Useful but shallow.
Tier 3: Harness users. Running OpenClaw, Cursor, or other tools with Claude under the hood. Real leverage, but borrowed infrastructure at a subsidized price. When the price changes, the workflow breaks.
Tier 4: Builders. Working in Claude Code, in the terminal. They have subject matter expertise and they’re building things that compound. Their workflows survive pricing changes because they understand what’s happening under the hood.
The post argues that the Tier 3 → 4 gap is where careers diverge.
I agree. But the framework is missing a tier.
Tier 5: System Builders
Tier 4 builds workflows. Tier 5 builds systems that encode judgment.
Here’s the difference.
A Tier 4 developer writes a Claude Code script that:
Pulls Google Ads metrics from Databox
Formats them into a table
Adds a summary paragraph
That’s a workflow. It works. It saves time.
A Tier 5 practitioner builds a system where:
14 agents know which data source has a 48-hour sync lag and checks the lag BEFORE flagging an anomaly
A confidence engine scores every finding on a 0-1 scale based on source reliability, evidence classification, and assumption verification status
When Google Ads says “conversions dropped 74%,” the system generates 6 competing hypotheses, scores each against the evidence, and only creates a task if the confidence is above 0.4
The system knows that Google’s “conversions” use a 30-day click window while Meta’s use 7-day click — and flags the mismatch when someone tries to compare them
A recommendation tracker checks whether the advice it gave last quarter actually moved the metric — and if it didn’t, says so before recommending the same thing again
Per-client alarm calibration means a 15% drop is noise for one client and a crisis for another — because the system knows each client’s baseline and volatility
That’s not a workflow. That’s judgment, systematized.
Why the Gap Between Tier 4 and Tier 5 Is So Hard to Cross
Tier 3 → 4 is a technical gap. You learn Claude Code, you understand the terminal, you build things. Effort + learning = progress.
Tier 4 → 5 is a domain expertise gap. No amount of technical skill will get you there without subject matter knowledge worth systematizing.
You need to know:
Which questions to ask (not which prompts to write)
What “normal” looks like (not just what “good” looks like)
When the data is lying (not just what the data says)
Which recommendation will actually work for THIS client, not just what works in general
This takes years. Not of AI experience — of domain experience. The 22-year marketing practitioner who learns Claude Code in a weekend is closer to Tier 5 than the developer who has used Claude Code for a year but has never run a marketing campaign.
The Proof
We open-sourced our system this week: github.com/arcanianHQ/arcanian-os
160 files. Not code — methodology. Skills, agents, SOPs, guardrails. The thinking that makes the tools useful.
Anyone can clone it. The files are right there.
But the files aren’t the system. The system is what happens when 22 years of diagnostic experience gets encoded into guardrails that stop the AI from being confidently wrong.
A confidence engine doesn’t come from reading the Claude Code docs. It comes from getting burned — from the first version of our system that flagged “anomalies” that were just sync delays. From the analysis that treated a narrative repeated in five documents as “evidence” when it was just a story. From the recommendation that looked smart on paper but didn’t move the number.
Each guardrail is a scar. Each rule is a lesson. The technical implementation took weeks. The judgment behind it took years.
What This Means for You
If you’re at Tier 1 or 2: start using AI for real work. Today. The cost of watching is higher than the cost of trying.
If you’re at Tier 3: the author of the original post is right. The convenience you’re borrowing has an expiration date. Learn what’s happening under the hood.
If you’re at Tier 4: the question is no longer “can I build this?” It’s “do I have something worth building?” The most powerful Claude Code workflow in the world, pointed at the wrong problem, is just a faster way to waste time.
If you want to reach Tier 5: stop optimizing your tools and start systematizing your judgment. What do you know — from experience, from failure, from years of practice — that nobody wrote down? That’s what belongs in your CLAUDE.md. Not another prompt template. Your actual expertise, encoded in rules the AI can follow.
The files are open: github.com/arcanianHQ/arcanian-os
Not because the files are the value. Because the approach is
.