Lesson 5 of 5
How do I know the plan works before I need it?
A recovery plan written at a desk has not recovered anything. This lesson writes the operating plan for the scheduled report, says how each line of it can be checked by rehearsing the failure it names in a test setup that touches nothing real, reads two guides to the same product that disagree to show how setup documentation has to be read, and closes with a transfer task and the question that returns the reader to the setup decision record.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
The operating plan
An records, for one workflow, where it executes, what access it has, how progress is saved, how it is supervised, what failures are expected and what the response to each is. Here is the treasurer's, written for the scheduled report:
Operating plan: supplier invoice report, scheduled
Execution: harness process on the rented host, started on the 1st inside a
tmux session named "report"; the model runs at the provider; the
phone is a view that can answer asks and send text
Access: mailbox and board credentials in the host's credential file, each
service authorizing its own; host login, viewing-service login and
the credential file all protected; harness rules on the host:
reads free, write and post ask; steps run one at a time
Asks: if no answer within 30 minutes, stop and report; do not proceed
Saved progress: progress note updated only when a step completes, written to a
replacement file on the same file system, then atomically replaced;
storage-specific durability steps checked for the planned failures;
the report written the same way
Supervision: attach to tmux from any device; completion notification carries the
post receipt's item id and a hash of the report file
Storage: the host's disk is assumed to survive a restart; nothing restarts
the report automatically
Failures and responses:
client disconnect: reconnect and read; no action
process exit: read the note, the folder and the board; match the board
item's contents or receipt to the saved report before
marking the post done; restart only with completed steps
marked
host restart: confirm from uptime; no tmux server; read stored work; set
aside a partial report and rewrite from inputs; read a
partial note as evidence to check; check the board before
posting
missing notification: inspect the host; silence settles nothing
Unknown post outcome: as module 8: inspect the board; finish on a positive
finding; stop and report on an unresolved absence
The plan reuses module 8's recovery plan for the post and adds the location-dependent lines: where the process and the model run, where each credential lives and who authorizes it, what to do at an ask with nobody there, what the case assumes about storage and restarts, and what each of the four cases leaves behind.
Checking the plan in a test setup
Each failure line names an event and a response, and each can be rehearsed on purpose, but only in a setup built for it: a copy of a month's inputs in a separate test folder, a test mailbox and a test board that the committee never sees, and either a host dedicated to testing or a controlled simulation on the real host that does not touch the real run. Nobody restarts the production host to see what happens, and nobody sends a real report to the committee to learn whether the plan works. With that setup: disconnect the phone during a test run and reattach, and see whether the run continued. End the test harness deliberately after a post to the test board and start a fresh run, and see whether it matches the item and refrains from posting. Simulate a restart by ending the tmux server and the harness together, remembering that this is not the same as a host restart, since the host's own startup and its disk are not exercised, and see whether the fresh run sets aside the partial file. Silence the phone's notifications and see whether the plan's inspection step finds the finished run.
Each rehearsal is a representative case in module 8's sense, with a known correct response, and each passing rehearsal is evidence about that failure, on that host, with that harness version and that month's inputs, and no wider. A rehearsal has its own limits to set: scripted, so that it does the same thing each time; run under the harness's test permissions rather than the production ones; and budgeted, so that a rehearsal that goes wrong cannot spend a month's allowance. This course performed none of these rehearsals. No remote setup was installed or run in preparing this module, and the state records in the previous lesson are authored. A reader who adopts the plan checks it in their own test setup, and the plan is not checked until they have.
Remote Control also documents version-dependent session resumption. 2 A plan using that feature must name and test its installed version, including whether recovery avoids repeating a completed post.
Check what the installed version actually loads
Setup guides are part of the recovery plan. Check which version and mode a guide describes, then test the behavior that the plan depends on. When guides disagree, the product's reference documentation is a useful starting point, but a test of the installed version provides stronger evidence about that setup. Two documents from one publisher are not independent confirmation, even when they agree.
Anthropic's reference documentation and training lesson, rechecked on 21 November 2026, now agree on a distinction that earlier versions described differently. Running Claude Code without an interactive terminal normally discovers the same local hooks, skills, plugins, tool servers and project instructions as an interactive session. A separate option, called bare mode, skips that automatic discovery. Bare mode also uses a different authentication path: it does not use the subscription login. Choosing it therefore changes both what context loads and how access is billed. 6 7
For the treasurer's report, the practical question is whether a restarted process receives the instructions and connections the report needs. Record the mode and version, and check the loaded context in a test run. An option that skips automatic discovery can remove instructions the recovery plan relied on. It also does not remove every setting: the reference documentation says bare mode still reads settings files and permission rules, and describes exceptions for explicitly supplied directories. 6
A more repeatable setup does not guarantee identical output. The runtime, environment, model version and inputs can still differ. Even with those held constant, generation can produce different wording. The report's acceptance criteria from module 8 decide whether the result is usable: correct totals, traceable invoices and no duplicate post. Matching the previous run word for word is not one of those criteria.
Transfer checkpoint: a newsletter build on a home desktop
Try the module's distinctions on a different arrangement. Everything here is authored.
The garden's monthly newsletter is assembled by an agent running on the editor's desktop computer at home, started on Sunday evening inside a multiplexer session. It reads contributors' files from a shared folder, writes the newsletter, and emails it to the members list from the editor's mail account, which is logged in on that desktop. The editor supervises from a laptop at work on Monday, over a remote connection, and in this arrangement the laptop can only view and send text. Nothing on the desktop restarts programs at startup, and the product in use has no feature that resumes a crashed session on receiving a message.
On Monday the laptop shows nothing new. The editor connects to the desktop and finds: the desktop's uptime shows no restart since Friday; the multiplexer session present with a shell prompt in the pane where the agent was; no agent process in the process list; a newsletter file in the folder, complete; the mail account's sent folder showing the newsletter, with the newsletter's text, sent to the members list address at 22:41 on Sunday; and a progress note reading "newsletter written; send: pending".
Before reading the guide, answer five questions. First, draw the five things for this arrangement and say which of them the laptop is. Second, which of the four cases happened, and what evidence says so? Third, which credentials were used to send, and where do they live? Fourth, if the editor sends the agent "don't send yet, I want to check it" from the laptop, what does that reach? Fifth, what should a fresh run do first, and how would that line of the plan be checked before next month?
Self-review guide. On the drawing: the laptop is the viewing device, the remote connection is the connection, the agent is the process, the desktop is the host and execution location, and the shared folder, the newsletter file, the note and the mail account's sent folder are stored work, the last of them held by the mail service rather than the desktop.
On the case: a process exit. The uptime rules out a host restart; the prompt in the pane and the empty process list show the agent ended; the complete newsletter and the sent-folder entry, whose text matches the file and whose recipient and time fit the run, place the exit after the send and before the note's update. The records do not say why the agent exited, and an answer that claims a cause has gone beyond them.
On credentials: the mail was sent with the editor's account, logged in on the desktop, so that credential lives on the host, and anyone who can act on that desktop as the editor can send as the editor; the laptop held only its own login to the desktop. The sent folder is the mail provider's record of sending, not of delivery to any member.
On the message: in this arrangement it reaches nothing, because the process has exited and the product does not resume a session on receiving a message. Had the agent been alive, the message would have reached only the run's future; the send at 22:41 is done, and a product with a resume feature might have restarted the run and acted on the message next, but could not have unsent the mail.
On the fresh run: it reads the note, the folder and the sent folder first, matches the sent entry to the newsletter by text, recipient and time, marks the send done, and does not resend. That line is checked before next month by rehearsing it against a test address from a test folder: end the agent deliberately after a send to the test address and confirm that the fresh run finds the entry and stops. An answer that says the multiplexer should have kept the agent running has confused a client disconnect with a process exit.
Back to the setup decision record
This module added a multiplexer, a remote view, a plan and a set of rehearsals to a workflow that module 8 had already made checkable and module 9 had shown how to split. In this authored case each addition answers a failure the case describes: the tunnel, the afternoon exit, the maintenance restart, the silent phone. Nothing here was demonstrated on a real setup, and a reader's own record will hold their own cases. A technique does not have to wait for a failure to have happened: a credible failure, one the drawing makes plain could occur, is reason enough to plan for it, and the plan's rehearsal is what turns the plan from a guess into evidence.
That is the test module 6's setup decision record puts at the center of any setup choice: a technique earns its place by addressing a problem that was observed or credibly expected, and it is retired when the assumptions it rested on no longer hold. The multiplexer rests on the harness needing a terminal that outlives the client. A product that reconnects on its own and runs without a terminal changes that line of the plan, and the host still has to stay up, since nothing about the view changes where the process runs. The record is where such a change is made and dated, with the rehearsal that showed it works. The plan is current for a host, a harness version and a month, and the person who runs it knows which.
A reader finishing this path has, from module 8, criteria and a recovery plan for one workflow; from module 9, a map and a plan for splitting it; and from here, an operating plan for running it elsewhere. The next decision is theirs: which of their own workflows is worth this treatment first, and what its first rehearsal should be.
Carry this forward
- An operating plan records execution location, access and who authorizes it, what happens at an ask with nobody there, how progress is saved, how the run is supervised, what the case assumes about storage and restarts, and the response to each expected failure.
- Each failure line is checked by rehearsing the failure in a test setup: copied inputs, a test folder, test destinations, a dedicated test host or a controlled simulation, scripted and budgeted. A passing rehearsal is evidence for that case only. This course rehearsed none of them.
- When two documents about one product disagree, prefer the implementation's description for your version, prefer checked behavior over both, and record the disagreement. Skipping local discovery reduces differences between setups; it does not make them identical, and identical setups do not guarantee identical output.
- A technique earns its place by addressing an observed or credibly expected failure, and is retired when its assumptions no longer hold; the setup decision record is where that is dated.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
You have reached the end of module 10 and of the course's advanced path. The module quiz below checks how these ideas fit together.
Module 10 quiz: put the ideas together
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
Sources · Reviewed 2026-09-29
- 2. Anthropic, Claude Code documentation: Continue local sessions from any device with Remote Control
- 6. Anthropic, Claude Code documentation: Run Claude Code programmatically
- 7. Anthropic Academy: Routines and headless (Claude Code in Action)
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.