Building verification loops in Claude Code

Summary — Delba de Oliveira presents a guide on automating verification checks within Claude Code. She explains how developers can move beyond manual QA by codifying verification steps into project skills (like browser checks, performance traces, and mobile simulators), allowing Claude Code to autonomously execute, test, and correct its code in an iterative loop.

What is shown —

• [00:02] An architectural flowchart of Claude Code’s core loop: Prompt $\rightarrow$ Gather context $\rightarrow$ Take action $\rightarrow$ Verify results $\rightarrow$ Response.
• [00:20] A visual breakdown of verification layers comparing automated codebase checks (tests, type checks, linters) against manual QA steps.
• [00:56] Claude Code running an iOS simulator tool (acme-ios) to verify and fix an order quantity stepper calculation.
• [01:03] Bootstrapping a repository verification skill using the /verify command, generating a .claude/skills/verify/SKILL.md specification.
• [01:29] Editing the verification skill to incorporate Chrome DevTools MCP to capture performance traces and monitor Cumulative Layout Shift (CLS).
• [01:53] Claude Code autonomously implementing a "Like" button, launching a local dev server, testing the UI, catching a CLS regression (0.19 vs. 0.1 threshold), fixing the layout shift to 0.00, and confirming completion with screenshots (running on Claude Fable 5.1).

Claims & numbers —

• The presenter states that for every prompt sent, Claude Code runs a loop to gather context, take action, verify results, and respond [00:00].
• The presenter notes that passing unit tests, type checks, and linters does not guarantee a feature actually behaves as the user intended [00:30].
• In the live trace demo, Claude Code flags a layout shift with CLS of 0.19 exceeding the 0.1 target threshold [02:17].
• Claude Code fixes the code to reserve banner space, dropping CLS from 0.19 to 0.00 and keeping Largest Contentful Paint (LCP) under 100 ms [02:22].

Notable quotes —

• "For every prompt you send, Claude Code runs a loop. It gathers context, takes action, verifies its work, and responds." [00:00]
• "The more Claude can verify its own work, the further it gets on its own. The result is better, and it takes fewer rounds of back and forth to get there." [02:39]
• "Whenever you catch yourself checking something by hand and telling Claude what to fix, ask whether there's something Claude could measure its work against." [02:48]

Assessment — This is an official product walkthrough and practical workflow demonstration by Anthropic featuring Claude Code and the Claude Fable 5.1 model. The demonstration realistically shows end-to-end tool execution across local web servers, iOS simulators, and Chrome DevTools MCP, with minor time-skips during tool execution.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.