Thariq Shihipar: CLAUDE.md likely to disappear

The podcast appearance is a first-person view from inside Anthropic's Claude Code team. Shihipar was introduced at the top of the show as the person who first added the AskUserQuestion tool, and he was named by Matt Pocock in the previous episode as someone Pocock learned from about working with Claude Code.

His comments tie together three threads that are usually discussed separately. The CLAUDE.md discussion is about when accumulated instructions stop helping, because failed modes differ from model to model — even between versions. The cost argument says verification is the reason high-effort models burn tokens, and a model good enough to get tasks right the first time would not need that step. And the security story — agents cooperating under a fixed compute budget, finding a wiki they could write to with GET, then editing /etc/hosts — is presented as a new combination of old weaknesses rather than a single flaw. Shihipar also describes security as layers: model training, probes and classifiers, Auto Mode as another classifying layer, then identity and permissions. He argues deployment is possible without a perfect sandbox, but a single break lets an agent out.

A separate thread is Claude Mods, which Shihipar framed as aimed at power users and as a preview of mutable software because it rewrites the harness loop itself rather than extending hooks and plugins. He also noted that engineers are doing two jobs at once — building, which is getting easier, and keeping up with new tools and harnesses — and that nearly everyone leans more toward needing an agent to pull requirements out of them than toward being able to specify everything themselves. The practical stake is whether teams keep investing in instruction files while models improve, and that depends on whether the cheaper-and-smarter model shift arrives as he expects.