In one sentence: Anthropic now has Claude run four automatic checks on its code before delivering it — and the competitive focus of AI coding is rotating from “who writes fastest” to “who verifies most accurately.”
What Happened
On July 22, the Claude Code team published its internal verification loop: after writing code, Claude no longer hands over the result directly. It runs four Skills in sequence —
-
/code-review — surfaces potential bugs and produces review comments;
-
/simplify — strips redundant implementations and simplifies structure;
-
/verify — runs the thing end to end to confirm the feature genuinely works;
-
/design — if UI changed, checks the visual implementation against DESIGN.md item by item.
All four pass, then it ships. Put differently: AI has graduated from “can write code” to “can check its own code.”
The Deep Change: the Bottleneck Moved
Here is the part worth sitting with. For decades, nearly every process in software engineering — requirements documents, planning cycles, layered reviews, endless meetings — existed for one underlying reason: writing code was slow, and engineers’ time was expensive. All that ceremony was scaffolding around a scarce resource.
When AI makes writing code fast and cheap, the premise evaporates. The Claude Code team’s own diagnosis:
The bottleneck didn’t disappear — it moved. From writing code to verification, code review, and security.
Code now generates faster than anyone can absorb it, and the new questions are: is this code actually right? Who maintains it? Can human reviewers keep pace with what the machine produces? This is not a tool update. It is a structural migration of where the AI coding race is run.
There is a second-order effect that makes verification self-reinforcing rather than merely important. Long-horizon agent runs — the kind where an agent works for hours without supervision — are only possible when the agent can trust its own intermediate output. Every check the machine performs on itself buys more autonomous runtime; every check that still needs a human caps the run at that human’s attention. Verification is therefore not just the new bottleneck for quality. It is the gating resource for how far agents can go at all.
The Architecture: Four Levels of Automation
Anthropic grades verification triggers from loose to tight:
-
Standalone — invoked manually, whenever you remember;
-
Embedded — wired into a task flow so it runs along;
-
Chained — several verification Skills linked to execute automatically in sequence;
-
On every PR — every submission passes through automatically.
The official name for the middle transition is “from habit to contract”: “I remember to run /verify after /simplify” is a personal habit; chained, it hardens into a system-level invariant. The chain walks the development loop on its own and comes back only for genuine decisions. Longer chains raise reliability — with the official caveat that chained verification burns real tokens, so prove stability first and escalate gradually.
A Skill Is Not a Prompt — It’s a Capability Module
One misconception needs correcting: a Skill is not a slab of Markdown prompt. It is a capability module containing instructions, file structures, scripts, tool calls, configuration, and an entire workflow — your team’s checks, design standards, and hard-won scar tissue packaged into something on call.
More consequential: Skills are becoming a cross-vendor open standard rather than a Claude Code feature. GitHub Copilot, Cursor, OpenAI Codex, and Gemini CLI have adopted the same format. The verification assets your team builds will not be locked to any single tool — an unusual and consequential property in this market.
The Consequence: Same Model, Several-Fold Efficiency Gaps
Which produces the uncomfortable reality: the same Claude, in different teams, delivers efficiency that differs by multiples. The gap is not the model. It is the workflow, and it reduces to three questions: Did you write your checks as Skills? Did you build a verification loop? Did you let the agent close its own feedback circuit?
An agent’s capability is an addition problem: model + tools + verification mechanism + workflow. Every vendor’s model term converges toward the same value. The other three terms are entirely in users’ hands — and they are where the distance opens up.
Action: Write Your First Verification Skill
-
Write down, in plain language, the step you perform manually every time — as if briefing a new hire on day one.
-
If you can’t articulate it, have Claude draft a generic best-practice version first, then edit. The points where yours diverges from the generic version are precisely the ones worth recording.
-
Checks need not be fuzzy judgments. “Any change that drops a database field without a matching data migration gets rejected” is a house rule no generic linter will ever catch.
-
Any red line you hold only by staring manually is a candidate for a loop.
Then what? Hand your notes to skill-creator and let it interview you into shape, or simply drop a Markdown file into .claude/skills/ — the simplest verification Skill is a few lines of instructions plus a body.
Editorial Judgment
To be clear about what this is not: this is not “AI now writes software independently.” It still needs engineers, and production-grade delivery still runs through people. But the direction is unmistakable. We spent decades teaching AI to write code; we are now starting to teach it to check whether what it wrote is right.
The day “one last manual review before clocking out” can be handed over with real confidence is the day AI starts carrying real weight for you.
