Queue New Ideas for AI as Reserved Work── Mark the request as one to start later and register only its content and done condition, and the running job keeps going while the idea is kept

01A new request sent mid-task blurs the line between the current job and the next

Hand an AI a large piece of work and it may run for tens of minutes or for hours. In that time, another idea arrives. The person using the AI now has a choice to make: send the idea at once, or wait until the current work ends.

Send it at once, and the AI may read the new request as an extra instruction for the job in progress. Wait, and the idea has to be carried in your head. Both choices cost something.

The operator's record shows one answer to this choice. In September 2026 the operator started opening requests with the label "reserved work" (in Japanese, sagyō yoyaku), and used it 17 times within the month.

The earliest reservation in the material, on 8 September, asked for a system that publishes Japanese and English articles at a fixed time every day. Four more followed on the 9th: fixing inconsistent page designs, keeping the update history and the two languages in line, adding a new section to the regulation guides, and a nightly search for mismatches between the Japanese and English pages. On the 10th came a screen layout for mobile phones, and on the 11th a request to strengthen the AI that writes the articles.

02The label sets the start time, not the content of the task

The label adds nothing to the content of the request. It adds one piece of information: when the work may begin. A labelled request means "leave the current job as it is, and take this up when that job is finished."

None of the reserved requests in the record were small chores. They asked the AI to build a system, create a new section, or set up a check that runs every night: work spread over several days. The larger the request, the more likely the operator was to label it and queue it.

Most reservations also said where the work ends, not only what to make. Fix the mismatches until none remain; run the check as a daily routine. The start was deferred, but the finished state was fixed in advance.

03Holding an idea back and pushing it in both carry a cost

Waiting costs the person. The idea has to be remembered, and if it is forgotten, the request is lost. The longer the AI's job runs, the more ideas pile up in the meantime.

Interrupting costs the AI's work. When a separate topic arrives in the middle of a job, it becomes less clear where the current job ends. If the job is stopped and switched, its steps and the decisions made so far have to be rebuilt later.

A labelled reservation reduces both costs at once. The idea goes to the AI as soon as it appears, so nobody has to hold it. Because the start is set for later, the current job keeps its shape.

The gap grows with the number of ideas. The operator queued at least seven requests over four days in September. When an AI's jobs run longer, more ideas arrive while you wait.

04Reserved work is a request that defers its start and states its finished form

This series defines reserved work as follows: a request that says the current job should not change and that work may start later, and that states its content and its condition for completion. One message carries four parts.

The boundary matters too. A request that changes the way the current job is being done is not a reservation. It is an interruption, and it should reach the AI first, even if the job has to stop. Episode 8 covers how to hand one over.

This episode also leaves aside which reservation to start first when several have piled up. Episode 6 covers setting the order yourself, and Episode 9 covers making sorting and priority part of the AI's own system. The subject here is only this: registering an idea on the spot without changing the job in progress.

05In material review, one review runs to the end while improvements accumulate

Pharmaceutical material review is work in which switching midway invites mistakes. When a reviewer is checking the claims in a product brochure against the package insert one by one, a break for something else makes it unclear how far the check got.

The same holds when an AI does the first pass. While the AI works through every claim in a piece, the reviewer thinks of improvements. Count past review comments by type. Run the same check on the English version of the material. Ideas like these do not change the check that is running.

Queued with a label, they leave the running check on one standard from start to finish. If the standard shifted midway, the first and second halves of one record would rest on different checks. That mixture is what makes a review record hard to explain later.

The same applies in medical affairs and marketing. While the AI summarises a set of papers, you think of a structure for a briefing deck. While it organises competitor materials, you think of a format for an internal table. Each can be registered as the next job without touching the current one.

06People are distracted by unfinished tasks, and AI loses track in mixed context

Start with the person. In a 2011 psychology study, unfinished goals led to intrusive thoughts about them during an unrelated reading task and to weaker performance on a separate task. When participants made specific plans for those goals, the effects disappeared.

That result matches what a reservation does. A reservation writes an unfinished task down in the form of when, what and how far. Once it is written down, attention returns more easily to the current job.

The cost of interruption has been measured as well. In an experiment published in 2008, people who were interrupted finished their work in less time than those who were not, with no loss of quality. But they reported a higher workload, more stress, more frustration and more time pressure. The speed was paid for by the person.

Now the AI. Anthropic's guidance for Claude Code warns that moving unrelated tasks in and out of one conversation fills the context with irrelevant information, and recommends clearing the context between unrelated tasks.

In the same Claude Code, a message typed while the AI is working is queued and passed to the AI as soon as its current tool calls finish, inside the same turn. A request sent mid-task, in other words, can be read as part of the job in progress. The label tells the AI not to treat it as an instruction for that job.

07From tomorrow: label, content, done condition and start condition in one message

There are four steps.

1. Open with the label. Write "reserved work" and add that the current job should not change.
2. State the content and the done condition. Say what, for what purpose, and where the work ends, in one or two sentences.
3. State the start condition. After the current job ends, or after your approval.
4. Have it written to a list and read back. Ask the AI to keep reservations on a list outside the conversation, and to list the queued reservations at the end of its report on the current job.

The fourth step has a reason. When a conversation grows long, Claude Code automatically summarises older exchanges to make room. If a reservation is dropped in that summary, a request you thought was queued is gone. A list kept outside the conversation survives the summary.

When reservations are listed at the end of each report, you can catch a missing or mistaken one on the spot. Deciding which of them to start next stays with you.