2026.09.27 2026.09.27

How Far You Can Step Away — A Reverse Guide to Running Claude Code by Cost and Recovery

swiftwand

Six articles in this series each looked at one surface. This one puts them side by side, because the question that actually comes up is not “what can a routine do” but “I want to do this thing, where should it run, and what happens if it goes wrong”.

Two axes decide almost everything: whether a human is there to press approve, and how expensive the mistake is to undo. Everything below is an arrangement of those two.

Six surfaces, four questions
Three questions, in order
The reverse table: start from what you want to do
Delegating and stepping away are two decisions
Cost, in one paragraph each
The arrangement at this bench
Read the rest of the series
The short version
Sources

忍者AdMax

Six surfaces, four questions

Surface	Machine on?	Who approves	What leaves your machine	Minimum interval or scale
/loop, in-session scheduling	Yes, and the session too	Inherits from the session	Same as the session	1 minute. Expires after 7 days, 50 tasks per session
Desktop scheduled task	Yes	Configurable per task	Stays local	1 minute
Routine (cloud)	No	Nobody	A repository clone; connector actions appear as you	1 hour, plus a daily run cap that varies by plan
Workflow (local)	Yes	Confirmed at launch, mode-dependent. Auto asks once, bypass and -p never ask. Agents follow your permission rules	Same as the session	16 concurrent agents, 1,000 per run
Remote Control	Yes, it runs locally	You, from a device	Conversation records on Anthropic servers	Not applicable
Artifacts	Only while building	Publishing depends on permission mode	Private by default; a public link is readable by anyone	One page, 16 MiB

The second column is the one people skip and then regret. Only the routine row says No, and everything else about routines follows from that single fact: no local files, no approvals, a one-hour floor, and a daily cap.

Three questions, in order

Does it need a file that is not in git? If yes, it cannot be a routine. A routine sees a fresh clone and nothing else.
Does it need the machine to be awake? If you cannot guarantee that, it has to be a routine. Desktop tasks skip runs when the computer sleeps, and /loop needs an open session on top of that.
Does its last step change something in the world? If yes, it needs a human in the chair, which rules out the routine and points at a local session, a workflow with approvals, or Remote Control with push notifications turned on.

Answer those three in order and the surface usually picks itself. The cases where it does not are the interesting ones, and they almost always mean the task should be split.

The reverse table: start from what you want to do

What you want	Surface	How cost shows up	How to undo it
A weekly fixed check before publishing	Routine on a schedule	Subscription usage per run; the daily cap varies by plan	Read the run transcript. Work lands on claude/-prefixed branches and the default branch is untouched unless the prompt says otherwise
A fixed check that has to reach outside websites	A local session, or /loop	Same as the session	Our routine had 27 of these refused at the Trusted network boundary. Run it locally instead
The same inspection across dozens of files	Workflow	Proportional to agent count; an advisory warning past 25 agents or 1.5M projected tokens under the default guideline	Stop the run in /workflows. Completed agents keep their results, but a failed one reruns everything after it
Watching a long task from elsewhere	Remote Control	The same allowance as a normal session	Nothing to undo. The work never left your machine
Turning a formula into something a reader can use	Artifact	Output tokens; a styled page uses plenty	Republish. Each publish is a version and you choose which one viewers see
Reading a page you are already logged in to	Claude in Chrome	Browser tools loaded by default raise context usage	Nothing was changed, if you kept it to reading

Delegating and stepping away are two decisions

It is easy to collapse them. They are not the same. Delegating is about whether Claude can do the task well. Stepping away is about whether the task can survive being done wrongly with nobody watching.

A task can pass the first test and fail the second. Our source-link check is a good example: Claude handled it perfectly, and it still could not run unattended, because the environment it would have run in refuses the network calls it depends on. That is not a capability problem. It is a placement problem.

The reverse also happens. A task you would not trust Claude to finish alone can still be safe to leave running, if the worst outcome is a branch you do not merge and a report you ignore.

Cost, in one paragraph each

Routines draw subscription usage per run and have a separate daily run cap. One-off runs do not count against the cap.
Workflows scale with agent count, and the caches help: agents sharing a prefix read each other’s, though workflow agents get a five-minute cache TTL by default even on a subscription.
Everything on the main conversation gets a one-hour cache TTL on a subscription within plan usage, and drops to five minutes once usage credits are in play.
Fable 5.1 is half price on cache reads and double on everything else, and may bill to usage credits depending on your plan and seat tier.

The arrangement at this bench

For a solo operation running a blog and a small 3D printing practice, the split that has settled out is this.

A weekly cloud routine that reads the repository and reports, with its connector list emptied and its scope limited to the default branch
Local workflows for anything that repeats across many files, started at the desk and watched from a phone
Remote Control turned on for runs longer than about ten minutes, with push notifications for permission prompts
Artifacts for anything a reader would otherwise have to retype
Nothing at all on an unattended surface whose last step publishes, deletes, prices, or buys

The last line is the only rule in that list. The rest are preferences that will change as the tooling does.

Read the rest of the series

What Changed in Claude Fable 5.1, on the one row that moved in the price table
Leave Claude Code on Watch with Routines, on cloud scheduling and the network boundary that broke our link checker
Claude Code Workflows Pin the Script, on running dozens of subagents from one file
Drive Your Local Claude Code from Anywhere, on approving from a phone and where Chrome stops
Turn a Workshop Calculation into a URL with Artifacts, on the calculator and the comparison with ChatGPT Sites
Decide Claude Code Permissions and Cost, on where the money actually goes

The short version

Two axes decide the surface: who presses approve, and how expensive the undo is
Ask in order whether the task needs local files, whether it needs the machine awake, and whether its last step changes anything in the world
Routines are the only surface with nobody in the chair, and every other property of theirs follows from that
Delegating and stepping away are separate decisions; a task can pass one and fail the other
Put nothing whose last step is irreversible on a surface with no approvals