🧑‍💻 Caught lying to itself

WELCOME

gm builders, happy Wednesday. This week in four numbers, a prompt worth stealing before Claude changes again, and a forum thread that pulled 94 replies on what actually happens once your agent's mistake reaches someone.

THIS WEEK, IN NUMBERS

3 — how many times one agent's "posted" was a lie. The thread →

5 — skill files that just beat 300. The repo →

94 — replies to this week's sharpest AI safety debate. The thread →

½ — what Opus 5 costs against the flagship it nearly matches. The launch →

Five builds worth knowing

STEAL THIS

Turn 30 days of Claude Code sessions into a mirror

@PrajwalTomar_ built this as a self-audit, not a one-off. Point Claude Code at your own last 30 days of sessions and it turns whatever you kept doing by hand into an actual skill, and flags the biggest inefficiency in how you prompt:

FROM THE FORUMS

Your agent does something wrong. How far does it reach before anyone notices?

@aadilghani, who builds Pushary, asked builders where they actually draw the line for needing a human in the loop, and got 94 replies in a week. The best ones didn't agree with each other so much as sharpen each other. @raffay_sajjad, building a finance app, realized reversibility was the wrong thing to worry about: you can refund a charge, but you can't un-send the rejection email that went with it.

The sharpest reply, from @thenameisarian, reframed the whole question. Reach alone isn't the risk, reach without an accountable name attached to it is. Their fix is a compound gate: if an action's reach crosses a threshold and no named human attached their signature in the last few seconds, it pauses automatically. Not a policy written down somewhere, a mechanism that actually stops the action.

Vibebuilding

Every Monday

Tools, patterns, and shipped examples for builders using AI to make real products.