7/ new task = clean context
reusing stale instructions/sessions is how weird behavior starts
people call this "agent randomness" but it's usually state leakage
treat agents like junior operators: clear task, bounded access, clean handoffs
over the past few months i've been helping startups + solo operators roll out agents in real workflows
same pattern keeps showing up:
agents aren't bad at reasoning
humans are bad at operating them
most failures are due to context + workflow, not model failures
6/ fourth fix (big one): always run a two-step loop
step 1: propose plan
step 2: execute only after approval
this one change helps avoid a lot of expensive mistakes, especially for nontechnical users
Announcing a new Claude Code feature: Remote Control. It's rolling out now to Max users in research preview. Try it with /remote-control
Start local sessions from the terminal, then continue them from your phone. Take a walk, see the sun, walk your dog without losing your flow.
@KTmBoyle Spent two years in Miami, and this was truly one of my favorite aspects of the social scene. In SF, in most social settings, the default conversation always gravitates toward what you do for work, which gets boring quickly