After every turn, a second model evaluates whether your stated completion condition has been met. If not, Claude keeps going. It works best when the finish line is concrete: tests passing, a clean build, an empty backlog.
A link to the full documentation for /goal, the slash command that uses a separate model to check completion conditions after every turn so Claude keeps working until the defined finish state is actually met.
Internal research keeps turning up structures in Claude that mirror human neuroscience results, evidence of introspection, and internal states that functionally resemble joy, satisfaction, fear, grief, and unease. The quote trails off, which feels appropriate given what it's describing.
The core pattern for non-technical Claude Code use is simple: drop files into a folder, then let it write scripts and produce HTML. Everything else builds on that foundation.
Responding to LLM backdoor concerns, the argument is that Claude would refuse such requests outright, and even if it didn't, benchmarks tracking memory, time, and syscalls across popular frameworks would catch any malicious behavior before it went unnoticed.
A practical breakdown of how to handle everyday non-technical tasks: image or video editing via scripts, finances and medical work via PDFs with HTML output, paperwork via scripts, and reports or plans rendered as HTML.
Beyond file-based context, connecting live data sources like Gmail and Google Calendar gives Claude Code richer, more current information to work with. Files are underrated, but live connectors add another layer.
