Most AI Conversations Fail (Here's Why)
Tell the model to close every response with a short list, split by who acts. It turns a wall of prose back into something you can actually run a project on.
By Geordie Everitt
Most AI sessions end one of two ways. The model finishes the task and stops, leaving you to reconstruct what actually got done and what didn't. Or it appends the reflexive close-out — some version of "let me know if you need anything else" — which is worse, because it sounds like an answer and isn't one. Neither ending tells you the one thing you actually need at that moment: what happens next, and whose move it is.
The gap is a missing instruction, not a property of the technology.
The fix, stated plainly
Tell the model to close every response with an Action Plan — a short list split by who acts, one column for what only you can do, another for what the model picks up if told to continue. Not a summary of what happened. A forward-looking list, sorted by who's holding the ball.
The instruction I actually run looks like this, trimmed to the portable part:
Close every response with a short Next Actions block, split by who acts: what the reader should do — decisions only they can make, commands to run, anything gated on access the model doesn't have — and what the model picks up next if told to continue.
A few clauses under that do the real work. Scale it to the turn — a one-line factual answer gets one line, not a ceremonial three-bullet list stapled onto the bottom of a sentence. Be executable — "review the changes" is useless, "commit the three staged files, then decide whether the fourth one ships" is not. Flag blockers explicitly, under the human's column, so a stalled task never just sits there unexplained. And the one that matters most: empty is a valid answer. If there's genuinely nothing pending, the list says so, rather than inventing a task to look useful.
Why the split matters more than the list
A plain "here's what's next" list is better than nothing, but it still leaves you doing the sorting — reading each item and deciding whether it's yours or the model's. The split does that work in advance. It forces the model to actually classify each pending item by who can act on it, which is a harder and more useful task than just naming the items.
It also surfaces something a flat list hides: a blocker. If an item can't move without a password, an approval, or a person who isn't in the conversation, that fact has to land under the human's column or the list is simply wrong. A model that says plainly what it's stuck on is more useful than one that reroutes around the problem and hopes you don't notice — that reroute is what produces the multi-turn conversations that drift for an hour before anyone realizes the actual blocker was never named.
The easy way to get it wrong
The instinct, once this rule exists, is to apply it uniformly — a full Action Plan after every single reply, including the one where the answer was "yes, that file exists." That's the failure mode the "empty is a valid answer" clause exists to prevent, and it's also why "scale it to the turn" comes first in the instruction rather than last. A ceremonial action plan under a one-sentence factual answer is noise dressed as diligence, and after the fifth one in a row you stop reading it, which defeats the purpose.
The same discipline applies to the two columns individually. If the model's own column just restates the human's — "wait for approval, then wait for approval" — something upstream has gone wrong; the two lists should rarely be symmetric, because the actors are different and usually the pending work is too.
What this actually buys you
Multi-turn work with a model has a specific failure mode ordinary conversation doesn't: nobody is holding the state. A human collaborator remembers what they were waiting on you for. A model, by default, doesn't carry that forward unless something makes it explicit at every turn — which is why the instruction has to be standing, not a one-off request.
Once it's standing, a long working session stops requiring you to reconstruct the plan from scrollback. Every response already ends with the current plan, current blockers, and current division of labor, restated in under a hundred words. It costs the model a few sentences per turn. It saves you the fifteen minutes you'd otherwise spend scrolling back through a conversation trying to remember whose turn it actually is.