Detailed Analysis
A Reddit post detailing the construction of fail.ticker.io—an automated AI incident tracker—offers a granular window into how Claude Code is being used not just to write isolated scripts, but to architect and maintain a full production system with minimal human intervention. The site aggregates three disparate AI incident databases (the OECD's AI Incidents Monitor, the AI Incident Database, and MIT's AI Risk Repository) into a single daily-updated digest, solving a real annoyance: each source database has its own clunky interface that the developer compares to "filling in tax returns." What stands out is less the end product than the development process the author describes, which leaned on Claude Code for tasks spanning UI design exploration, reverse-engineering undocumented APIs, and building a dependency-free deployment pipeline.
Several details illustrate a maturing pattern of agentic coding practice. Rather than iterating on design through text descriptions, the developer had Claude Code build three complete, functioning design variants—a data-journalism layout, a mock newspaper broadsheet, and a dry CDC-report aesthetic—directly on the live site behind a toggle, letting the human evaluate real rendered pages with real data before choosing. This "show, don't describe" workflow reflects a broader shift in how developers are learning to collaborate with coding agents: instead of specifying requirements upfront and hoping for the best, they let the agent produce multiple concrete artifacts and make decisions empirically. Similarly, when the OECD's incident monitor turned out to have no public API, Claude Code was tasked with digging the internal endpoint out of the site's JavaScript bundle, discovering its 100-result pagination cap through trial and error, and then writing a windowed backfill strategy to work around it—the kind of exploratory, iterative debugging that previously required a human engineer's patience and web-inspector skills.
The infrastructure choices are notably minimalist and reveal thoughtful systems design baked into the agent-assisted build: a fully static site with no framework, no build step, and zero npm dependencies, deployed via versioned artifacts on Buddy CI/CD with no caching layer, so publishes go live instantly. Three scheduled pipelines—auto-publish, a daily "collect-incidents" workhorse, and a weekly AIID refresh—handle git-triggered deploys, data collection, editorial passes (again via a Claude Code action), page rendering, and validation. The most technically clever detail is the state-management trick: the daily pipeline's "memory" is the live published site itself, curling yesterday's JSON from production before fetching new data, explicitly to prevent a routine code push from ever accidentally reverting live data to stale repo contents. This decoupling of "state lives in prod, code lives in git" is a nontrivial architectural insight, and its presence in an agent-assisted side project suggests Claude Code is capable of contributing to real systems-design reasoning, not just boilerplate generation.
This example fits into a broader trend of developers using coding agents like Claude Code for end-to-end ownership of small-to-medium production systems—handling design experimentation, adversarial API discovery, CI/CD pipeline construction, and data-integrity safeguards within a single project, often with a human acting more as reviewer and decision-maker than line-by-line author. It also reflects growing interest in AI safety and governance tooling itself: an incident tracker that aggregates OECD, MIT, and AIID data addresses the fragmentation problem that has made it hard for researchers, journalists, and policymakers to get a unified view of real-world AI harms. That the tool was built largely through natural-language-directed agentic coding, including a self-described "FAILCON level" computed from live data, underscores how increasingly sophisticated, live-data-driven applications are becoming feasible for individual developers to build and maintain with AI assistance, compressing what would once have been a multi-week engineering effort into a rapid, iterative build process.
Read original article →