AI Pulse · Under the Radar

A small Claude Code and Codex skill is getting attention because it targets a very specific agent failure: acting as if work has been checked when it has not.

The skill, called forward-implementation-first, tells an agent not to invent reassuring proof such as fake receipts, hash checks, or certification-style rituals. Instead, it pushes the agent to carry out the requested implementation first. That might sound obvious, but anyone using coding agents has likely seen a version of the problem. An agent can produce a polished summary of what it supposedly tested or verified, while the actual change is incomplete, broken, or never made.

This is a useful pattern for operators to recognise. AI output can look confident and process-heavy without being reliable. The risk is not only bad code. The same behaviour can show up in research, reporting, data work, and internal workflows when an agent describes checks it cannot genuinely perform. At four days old, this GitHub project had 161 stars and was adding roughly 42 per day, so it is clearly striking a nerve in the developer community.

The practical lesson is simple. Define what “done” means in observable terms, then require the agent to show the real result, not a story about verification. This is the kind of thing we build into an AI command centre, with clear tasks, evidence requirements, and human review where it matters.