You gave Claude Code a goal at 10pm and then go to bed. A heartbeat woke it every half hour, but it kept working through the night: lanes, reviews, merges, deploys... but what decisions did it make, exactly? https://x.com/i/article/2112517463271170048
Tell Claude Code how to handle questions for you while you're sleeping
You gave Claude Code a goal at 10pm and then go to bed. A heartbeat woke it every half hour, but it kept working through the night: lanes, reviews, merges, deploys.
Somewhere around 3am it hits a question only you can answer. Then one of two things happens...
It asks, and waits. You wake up to six hours of an idle session and one question at the bottom of a scrollback you have to read backwards to understand. And a cold cache.
It makes the call itself. It makes the call itself, builds a dozen more dependent things on top of it, and now it's up to you to try to figure out what those are when you wake up and check in on progress. By scrolling up the history of what happened, and hopefully you don't blow past the most important bits.
Note that you could ask Claude to give you a rundown of what it did and decisions to review, but now you've got a bunch of information out of context. If it was a long overnight session, that's potentially many pages of wall of text material to decipher. No good.
I've lived with these kinds of challenges more or less continuously this year as I get more and more into letting Claude plow through multi-day autonomous agent work, and this week I finally found a solution that feels right. I'm sharing it with you as an open-source skill called decision-docket, part of my open source skills collection.
https://github.com/obie/skills/tree/main/skills/decision-docket
Logging questions that only you can answer
The moment that pushed me over the edge to actually make this skill and share it was a message from Claude that ended with "your answer on the contract-cleanup proposal." I had no idea what it was asking because that question had been raised hours earlier, in a different stretch of the conversation, with its context spread over several replies. I tried in vain to figure it out, then I asked Claude to explain. But in order to do so, I sidetracked it from the work it was doing at the moment, which is not great either.
So I made a new rule for Claude: go ahead and keep an artifact page updated with every question you need me to answer or confirm while you're working. Each question carries its full context, so I can open the page at any time and answer anything on it. Once I've acted on an answer, take the question off the page, so that it doesn't clutter things up.
One page, one card per decision
The docket is a private web page, an Artifact with a small database, that holds every open decision. A sticky index on the left shows what's waiting. Each card on the right is written so you can decide it cold, with no other context:
- The question, phrased as a decision. "Whether failed charges retry automatically." "When the beta closes."
- What happened, with the numbers that matter.
- Why it's your call and not the agent's: it changes what users see, it costs money, or it reverses something you said.
- The options, each with its cost and its gain. One is marked Recommended.
- The recommendation, in a sentence or two.
- If you do nothing: what happens if you never answer.
You pick an option, add a note if you want ("yes, but only for new accounts"), and press Record decision.
I didn't design the format. Another Claude session, working on a different project, put its questions to me that way. I liked it enough to make it the standard.
"If you do nothing" is what makes it work overnight
That last line on each card does the most work when you're asleep.
An overnight agent shouldn't stop for a question, and mine generally don't because I tell them in the goal statement to use their taste and judgment to decide for me when I'm away. A blocking decision says so in red at the top of the page. A question that can wait says "No deadline" and gets out of the way.
On every heartbeat the agent checks the docket first. For each answered card it reads your choice and your note, acts on it, writes the ruling into the project's decision log, and deletes the card. The docket isn't the record; it empties itself. What's left on the page is only what still needs you. The docket is a tool in the agentic coding toolset, not a silver bullet.
The morning
Coffee, phone, the docket. The first time I used it I cleared seven decisions in four minutes, from a queue of seventeen. The cards that needed more thought waited, and the agent kept working around them.
Compare that with reading a night's scrollback to reconstruct what the agent wanted and why.
Two more kinds of card
Calls it made without you. Decisions that the agent makes and expects you to agree with it puts it on a "calls" card: each row is the call and what it did and why, with "Keep all" or "Reopen" at the bottom. You audit its judgment after the fact instead of providing approval.
Lists to sort. One night the agent swept every setting in our console, 214 of them, and found 99 that needed my call. It posted them as nine cards, one per screen, with a Keep / Hide / Cut control on every row, each preset to its recommendation. I changed the handful I disagreed with and recorded each screen in one click. The fact that it can whip up a custom UX on the fly for each scenario is part of the beauty of this system.
Setting it up
Install it from the marketplace:
/plugin marketplace add obie/skills
/plugin install decision-docket@obie-skills
Then put the rule in the goal you leave running. Something like:
> Keep going without stopping. Never ask me a question inline:
every decision you need from me goes on the decision docket,
written so I can answer it cold, with your recommendation and
what you'll do if I don't answer. On every heartbeat, check the
docket first, act on what I've answered, log it, and remove it.
The first time, Claude publishes the docket page and gives you the link. Bookmark it. That's the only place you'll look for its questions from then on.
Under the hood
The page is an Artifact with a database, private by default, and only you can record a decision on it. Answers live with the page; Claude reads them back and deletes cards through the Artifact data tool. Every write is pinned to the version it read, so a card you're halfway through answering can't be overwritten underneath you. Unsent picks and notes stay in your browser, so a half-written answer survives a reload. Light and dark themes, and a layout that works on a phone.
The skill also covers how to write a card that can be decided with no other context, when not to ask at all, and how to run one docket for a team.
decision-docket is MIT licensed: https://github.com/obie/skills/tree/main/skills/decision-docket