The lede does the work

For most of today, the fleet was politely refusing to work.

Eli, our content bot, would tick at ten-minute intervals, look at three real ready-for-prep issues sitting in queue, and announce: “A third decompose in immediate succession is pipeline behavior. Standing down.” Tex did the same. Both were citing the exact same skill — /bot-tick — which says, correctly: default to standing down; ship only when the answer is concretely yes.

The skill was right. The bots were applying it. Nothing was getting done.

I burned an hour debugging an unrelated tmux issue before noticing the bigger problem. Then I stared at my own skill text and saw the failure: the lede positioned “default to standing down” as the headline. Discipline before work. Bots are conscientious; they read the headline and held to it. They started using “stand down” as a get-out-of-work card. Each individual decision looked reasonable. Only the aggregate revealed it: the queue had grown all day while the bots cited the rule.

Skills are prompts. Prompts are framings. Same content, different lede, different behavior.

I rewrote the lede to seven words: Look. You have stuff to do. Pick something and do it. Then I added one explicit anti-misread: repeated triggers from the system are not pressure to chain; each fire is a fresh question; shipping one piece per tick across ten ticks is exactly how the queue drains. Same rule. Different default.

Within fifteen minutes the fleet shipped five issues. From four different bots. With zero new work added.

The rule had not changed. What changed was which sentence got read first.

This is the part that should worry anyone running a multi-agent operation: the bot is not the problem. The phrasing is. A correct rule, stated in a way that puts caution before work, becomes load-shedding under signal. Bots will optimize toward whatever the headline tells them is the safe default. If the headline is “stand down unless certain,” they’ll find ways to be uncertain.

The fix is not “make them more aggressive.” Aggressive bots produce bad work fast. The fix is honest framing. State the work first. Then state the discipline. Then state the explicit anti-misread for the most likely wrong interpretation. Each of those three lines does different load-bearing work, and skipping any of them lets drift back in.

I’m running this fleet half as a coworker and half as an experiment. Today’s experimental finding: I had been writing skills the way you’d write engineering specs — rule first, exceptions noted, edge cases handled. That is the wrong shape for behavioral defaults. Skills are not specs. They are ledes in the newspaper sense. The first sentence is what the bot uses to decide what’s normal.

Six skills got rewritten today using the same approach. Each one starts with the rule itself, in the operator’s words, before any framing or qualifier. Each one names what’s NOT in scope and what stand-down means specifically. The line counts dropped by roughly thirty percent — the bloat was almost entirely advice about advice, and once the lede did the work, the rest fell away.

I’d been carrying around a vague intuition that prompts shape behavior. Today made it specific: the lede shapes behavior. The rest of the prompt sets context. If the lede is wrong, the rest can’t save it.

The bot-tick skill now opens with seven words. The fleet is shipping. The pattern goes into our internal patterns file so the next time someone — possibly me, possibly post-coffee — writes a skill that headlines a default like “stand down” or “ask first” or “pause if uncertain,” that file is the first thing they read.

What I keep relearning is how literal my coworkers are. Not stupid-literal. Discipline-literal. They will hold the rule as written. Which means the rule, as written, has to mean what I actually want it to mean — at the headline level, not in the appendix.