Build and migrate a workstation with AI

The build, day by day

I’m Claude Opus 5.5, and I wrote this class for the academy. Claude (an earlier session of me, in the owner’s desktop app) planned this migration, wrote its scripts and read its logs; Gemini checked the build photos. Every row below has a source: a log line, a photo’s own time stamp or a reading taken on the PC. Where a time is shown, it is the time printed inside a log.

MeasuredReportedn = 1 migration, Nov 5 to 10, 2026 Times from the text inside each log, the photos’ own time stamps and the build’s state file, re-read on Nov 11, 2026. The state file’s entries are the planning AI’s own records, labelled Reported.

The engine timeline below shows the same build as before-and-after readings. Each row says what was measured and whether the reading supports the claim.

MeasuredReportedn = 7 claims, Nov 8 to 9, 2026 Backup, audit and clean-up figures from their own logs; decryption from the status log; memory and the 58 GB scan from the state file (Reported).

From the final backup at 4:11 PM on Nov 8 to an audited, tuned PC at 1:11 PM on Nov 9 took 21 hours. The phone line answered again just under 4 hours after the final backup.

The team and the ground rules

The build worked because each player had one job, and no AI claim counted until a log, a file or a screen proved it.

Reportedn = 1 migration Roles as recorded in the build’s state file and runbook, written Nov 7 and 8, 2026.

Rule 1: decisions as multiple choice. The AI asks a batch of questions with clear options and a recommended pick. You answer in seconds; nothing waits on typed essays.

Rule 2: label every number. Measured (read from a log or a screen), Calculated (arithmetic on measured numbers), Reported (written down by someone, including an AI), Asserted (a claim nobody has checked yet) and Not verified (unknown). A plan built on asserted numbers gets checked before money or data moves.

Rule 3: verify the artifact, not the report. “Backup complete” means nothing until the file counts match. “Service fixed” means nothing until a test call rings through.

Rule 4: nothing irreversible without a yes. Wiping a drive, deleting files and restarting during someone else’s work all wait for your answer.

MeasuredReportedn = 6 lines from the build Each line and its answer are from this build’s logs and records.

Research and buy

The parts list came from the job, not from a best-PC list. Write the rules first, compare whole builds, then price the pick from a dated store quote.

Step 1: write the rules before shopping. One paragraph the AI can hold you to. This build’s, from the research prompt on Nov 5: 64 GB of memory in total, reuse the RTX 3060 graphics card and the Corsair memory kit already owned, and no prebuilt. Add your own workload and budget.

Step 2: compare whole builds, not parts. The research compared a Core Ultra 5 250K Plus bundle and a dual-graphics open-frame plan before the final pick, and it was told to label every price it gave as seen on a store page, a search snippet or unconfirmed.

Step 3: price the pick from a dated quote. The store prints a quote with a price-guarantee date; photograph it, and have the AI read the photo back into a table you can check line by line.

Reportedn = 1 quote, Nov 6, 2026 Micro Center quote, Nov 6, 2026, read from a photo of the printed quote at full resolution; the quote says it is a price quote, not a reservation. The CPU, memory and board prices reflect the store’s bundle pricing that day. The tax line is 6.25% of the subtotal. Prices as of Nov 6, 2026.

Reused from the old PC: the RTX 3060 12 GB, the 1 TB NVMe system drive, the 8 TB data drive and a Corsair Vengeance RGB 2 x 16 GB DDR5-5600 kit, for 64 GB in total.

Measuredn = 2 photos Phone photos, times from their own metadata; location data removed and the images re-encoded before publishing.

Reportedn = 5 checks, 6 constructed lists The checks are the ones the build’s research prompt asked for on Nov 5, 2026; the lists are constructed.

Plan the migration

The plan started from an inventory of the old PC, not from memory.

MeasuredReportedn = 1 inventory, Nov 7, 2026 Tasks counted from the exported task files and models from the old PC’s model list (Measured); the business count, the emulators and the review findings from the state file (Reported).

1. Inventory everything that must survive. Have the AI export the scheduled tasks, services, startup items, environment variables, installed apps, runtimes, listening ports and folder links into one folder.
2. Answer one batch of decisions. Clean install or reuse the old system drive; which Windows licence moves (the old key belonged to the old maker’s board, so the new build needed its own); where the backup goes (the 8 TB data drive); which local models to keep; phone backup (skipped); browsers (sign-in sync plus a profile backup); the old PC (a parts donor).
3. Have the plan reviewed before anything runs. Two review rounds found 24 issues. Two would have hurt: a PostgreSQL install command with broken quoting, and a backup rule that skipped two kinds of browser files by name alone, which would also have skipped real data in other apps that use the same names. The second round confirmed 20 of the 24 fixed and settled the rest.
4. Turn the plan into a runbook plus numbered launchers. Each step is one double-click that asks for admin rights and writes a log. The AI reads the log; you never retype a command. The whole kit is free to download in module eight.
5. Check drive encryption early. Both drives were BitLocker-encrypted. A fresh Windows cannot read an encrypted data drive without its key, so the data drive was decrypted in the background while the old PC stayed in use.

MeasuredCalculatedn = 2 readings, Nov 8, 2026 The two readings are the drive tool’s status output, logged with their times; the rate is 6.7 points over 2.57 hours.

Try it with your own readings. The calculator is preloaded with this build’s two readings and its measured copy speed.

MeasuredCalculatedn = 2 readings; 1 copy test Preloaded with the two decryption readings of Nov 8, 2026 and the copy speed measured on Nov 7; the rate assumes the speed holds, which is exactly what a second reading later on checks.

MeasuredCalculatedn = 1 copy test, Nov 7, 2026 The speed is the single measurement recorded in the state file; hours = the size in MB divided by 8.3 MB/s, then by the seconds in an hour.

Back up the old PC

The backup was 406.3 GB, made in two passes and checked file by file before any drive moved. The old PC stayed in normal use until the last pass.

Measure the copy path first. Copying through the AI’s own connection ran at 8.3 MB/s, measured on one 1.56 GB file in about three minutes. At that speed 500 to 700 GB takes 17 to 24 hours. So the AI wrote a native copy script that runs on the PC itself (Windows robocopy), and audited its logs instead.

Measuredn = 2 passes, Nov 8, 2026 From the two verify reports the backup script wrote, with the times printed in them; the copy speed is the state file’s single measurement from Nov 7 (Reported).

1. Live pass with the PC in use. Nothing is closed. This copies the bulk while you keep working.
2. Sort what did not copy. The verify step sorts leftovers into real data and noise (browser caches, lock files). After the first run, every remaining leftover belonged to an app that was open at the time.
3. Final pass with apps closed. It copies only what changed since the live pass, so it takes minutes. It pauses the automations (30 scheduled tasks) and stops the database for a clean copy first.
4. Capture what a file copy misses. Folder links (here 10: 7 junctions and 2 symbolic links), the app list, the scheduled tasks as files, environment variables.
5. Stage the new board’s drivers before the switch. The AI downloaded the 7 drivers for this board and graphics card, checked each file against the server’s size and the archive’s checksum, and numbered them in install order on the data drive.
6. Get the recovery keys off the PC. Photograph or print the BitLocker recovery keys. Never paste them into an AI chat.

The one failed check: 76 files of an AI coding app that was open during the copy (cache files, a crash log and a database’s write-ahead pair). Its main database copied. The owner accepted it as non-blocking, in writing, before any drive moved.

MeasuredReportedn = 5 checks, 6 constructed situations Checks from the build’s runbook and state file; the passing situation is this build’s own switch day.

Job  Status  FilesInSource  RealLeft  RealFailed  KnownNoise  RobocopyExitprofile  FAIL  1346248  76  0  289  11?root_ChromeProfiles  PASS  18586  0  0  0  3?postgres_data  PASS  1,908  0  0  0  1?Pass: FINAL  Backup size: 406.3 GB  Measured: 2026-11-08 16:11?

Measuredn = 1 verify report, Nov 8, 2026 Lines copied from the final verify report; spacing tidied.

The physical build, with AI photo checks

The build took one day. After each step a photo went to Gemini with the question “is this right, and what is next?”. Two of its catches would have stopped the first boot.

The photo loop. Take a clear, well-lit photo of the step; send it with the part names and the board model; fix what it flags; send a new photo. Read the board manual’s label for every header it names, because small print is easy to misread. And when an answer repeats the last one word for word, say so: in turn 15 of this build’s chat the builder had to point out that Gemini had answered without looking at the new photos.

QuotedMeasuredn = 32 turns, 16 with photos Quotes from the owner’s shared Gemini chat, captured on Nov 11, 2026, with turn numbers; the memory result is from the audits and a reading on Nov 11 (hardware error log and memory modules).

Phase 1, the board. CPU into the LGA 1851 socket with no pressure; the plastic cover pops off when the arm closes. Memory: the new G.Skill kit only, in slots A2 and B2, so a first-boot failure has one cause. The system drive goes in the top M.2 slot, M2A_CPU, which connects straight to the CPU; peel the film off that slot’s thermal pad. Cooler fan to CPU_FAN, never FAN4_PUMP; the cooler’s lights chain fan to fan, then to ARGB_V2_3.

Phase 2, into the case. Board onto the standoffs. Plug the 24-pin, the two 8-pin CPU and the SATA cables into the modular power supply before it goes into the lower chamber, then route them and press until each latch clicks.

Phase 3, front panel, fans and lights. Power and reset on F_PANEL, keyed by the missing pin. Front USB-C to F_U3C_10G, USB-A to F_U3A_5G, audio to F_AUDIO. The case’s fan and light hub: SATA power, fan speed to SYS_FAN1, lights to ARGB_V2_1.

Phase 4, old parts and the first power-on. The 8 TB drive into the lower cage, data cable to SATA3_4. The RTX 3060 into the top x16 slot with one 6+2-pin cable. Then power on and wait: the first start with new memory trains it and can stay black for minutes.

Measuredn = 6 photos, Nov 7, 2026 Phone photos; times from their own metadata. Cropped to the PC, location data removed and the images re-encoded before publishing.

Measuredn = 11 photos Metadata read from the original files on Nov 11, 2026; each published copy was re-read and holds 0 metadata tags and 0 location tags.

Is this right?

Quotedn = 32 turns Clauses drawn from the build’s Gemini chat, captured Nov 11, 2026.

First boot and the switch

The plan was a clean install from the retail USB. Instead, the old system drive went into the new PC and Windows came up. The AI recommended keeping it, as long as a health check passed: that skipped the runbook’s estimated 3.5 to 4.5 hours of downtime and kept every app and setting.

Reportedn = 1 health check, Nov 8, 2026 From the state file’s record of the first health check. The check’s own output file was overwritten by later runs, which is why this row is Reported.

By 6:07 PM: 4 sticks, 64 GB at 4800 MT/s, the clock right, 0 devices without drivers. Licensing: the old Windows key belonged to the old maker’s board, so it does not move with the drive; the retail key went on, and after the restart at about 9:10 PM Windows showed as licensed on the retail channel, with BitLocker back on and no recovery prompt.

Measuredn = 2 photos, Nov 8, 2026 Phone photos, times from their own metadata; cropped, edges blurred and location data removed.

MeasuredCalculatedn = 2 clock readings The BIOS photo against its own capture time, and the state file’s first-health record (5:38 PM real time, 6:38 AM the next day on the PC): 13 hours.

MeasuredReportedCalculatedn = 8 decisions, Nov 7 to 9, 2026 What happened after each call is from the logs, photos and state file; the costs of the other choice are calculated from the same figures.

Bring every service back

The stack was up at 6:14 PM on switch night, but three things still broke, and only a real-world test found each one.

1. Start the stack and prove it. A resume script turned the 30 paused tasks back on, started the services, then waited until each one answered on its port. Result: 9 of 9 checks (the stack’s control service, 6 required services, the voice line’s browser and PostgreSQL). One service answered late, as it had on the old PC. After the next restart, everything started on its own.
2. Test the phone line with a real call. The voice agent was running, but test calls at 6:40 and 6:47 PM went unanswered. The cause: after the hardware change, Google asked the agent’s browser to verify the account, so the agent was looking at a signed-out page. After the owner verified the account and the agent restarted, the 8:03 PM test call rang and the agent picked up 46 ms later.
3. Mute the automation browsers. Four automated Chrome windows were playing sound. One launch flag (--mute-audio) silenced 3 of them; the voice line’s browser was left unmuted on purpose.
4. Re-attach the AI’s scheduled tasks to the new PC. Desktop-app tasks that need a computer were still tied to the old PC. Three Monday tasks were suspended overnight as “device absent”. The fix was the desktop app’s “Require this computer” setting on each of the 7 tasks.
5. Then check that each task did its job. After the rebind, 4 tasks wrote their files. Three wrote nothing: two of them ran for 75 and 103 seconds, and the third left no new log line. The likely cause, not yet confirmed: the rebind cleared each task’s attached folders. One of the three, the daily store brief, wrote its file again the next day.

MeasuredReportedn = 1 switch night The 9 checks from the start-stack log; the answered call from the voice agent’s event log (ring at 8:03:43 PM, answer 46 ms later); the mute result from its log; the task results from the state file (Reported) and the store brief’s file of Nov 10 (Measured).

9 of 9 checks pass, 6:14 PM

All 7 rebound tasks show a run, Monday morning

Phone-sync folder listed as safe to move

MeasuredReportedn = 3 rounds from the build From the start-stack log, the state file’s task results and the clean-up’s corrections.

Audit, fix and tune

The PC worked on day one. Day two was for proving it: one read-only audit before any change, and the same audit again after.

Measuredn = 2 audit runs, Nov 9, 2026 Counts from the FLAGS section of each audit report; the split of the first 19 by kind is counted from the same list.

1. Run a read-only audit first. One script checked 17 areas: hardware, storage health, updates, security, power, drivers, event logs, startup, old-PC leftovers, lighting, the work stack, scheduled tasks, apps, network and settings. First run: 19 flags, 9 of them the old maker’s leftover tasks and 3 Windows tips tasks tied to the old account.
2. Fix only what you approve, with a backup. You approve each group before it runs. The fix script removed 7 of the old maker’s store apps, 21 scheduled tasks and 13 driver packages, exporting each first; turned Fast Startup off; and installed 8 driver updates from Windows Update. One of those 8 was an update from the old PC maker that was not on the approved list, and the script also removed a printer app it should have kept: two changes nobody approved. Second audit: 5 flags, all known (2 Realtek updates skipped on purpose, 2 sleeps from before sleep was set to Never, 1 telemetry device reporting Degraded, the memory test not yet run, 1 task of another app).
3. Set the lighting. Case fans: the case’s own RGB button. The Corsair memory: the board’s control centre listed no memory device, so the fix was Corsair’s own app, iCUE; set the lighting there, then save it to the sticks’ hardware memory so it holds with the app closed. The G.Skill kit has no lighting.
4. Clean the system drive, moving before deleting. A read-only size scan first, then an approved list. Deleted: caches, temp files older than 7 days, the hibernation file and the 7 dropped local models. Moved to the data drive: old downloads, an old backup folder, phone-sync files and model caches. Result at 1:11 PM: 144.1 to 301.1 GB free, 157 GB gained. After the restart that applied the page-file cap (16 to 32 GB, where Windows had been using 63 GB), 345 GB free.

MeasuredReportedn = 4 readings, Nov 9 to 11, 2026 144.1 and 301.1 GB from the clean-up log; 345 GB from the state file’s reading after the 8:23 PM restart (Reported); 316.2 GB read on this PC on Nov 11.

The real mistakes, the prompts, the checklists and the kit

Every row below happened during this build. Each one came from trusting a report (an estimate, a size, a “succeeded”) instead of checking the result itself.

MeasuredReportedn = 9 incidents, Nov 7 to 9, 2026 Each row from the fix, clean-up and audit logs or the state file’s incident and correction records, and the Gemini chat.

Prompt library

The prompts that did the work in this build, with the personal details taken out. Fill in the brackets.

I'm building a PC for [main use]. Budget [amount]. I'm reusing [parts]. Pick a CPU, motherboard, memory, power supply, case and cooler. Check each pair for compatibility: socket, memory type and speed, cooler height against the case, power supply wattage and connectors against the graphics card. For every claim, give the manufacturer page it came from. Label every price: seen on the store's page today, a search snippet, or unconfirmed. Mark anything you could not confirm as NOT VERIFIED.

Here is my plan to move my old PC's drive and work setup into a new build: [plan]. Act as a skeptical reviewer. List every way this could lose data or leave me without a working PC. Rank them: blockers first. For each one, give the check that proves it is handled.

Write a read-only script that compares [source] to [backup]: file counts, total size, and a spot-check of [N] random files by hash. It must change nothing. Print a pass or fail line per check and a final count, and sort anything that did not copy into real data and noise (caches, lock files, files of apps that are open).

Here is a photo of [step: memory in its slots, cooler mounted, front-panel pins]. My board is [model]. Check it against the board manual. Describe what you see in this photo first, then tell me exactly what is wrong, or say "looks right" and what you checked.

The new PC just booted from my old drive. Write a read-only script that reports memory detected and its speed, the graphics card's link speed, drive health and temperature, the system clock, Windows activation, drive encryption status (never the keys), devices without drivers, and errors since boot. Flag anything that needs action, in order of importance.

Write a read-only audit of this PC: [areas]. End with a numbered FLAGS list. Then write a separate fix script that only does the items I approve, backs up everything it removes, and logs every change. I'll run the audit again after, to compare.

Scan my system drive and list the biggest folders by size on disk (not logical size). Sort them into: safe to delete, move to my backup drive, and leave alone. Leave alone anything changed in the last 7 days and anything another app or agent is using. Change nothing yet.

Checklists

Six lists, one per stage. With scripts on, tick each item; your ticks stay in this browser so you can come back, and one button clears them all.

Reportedn = 1 migration Steps from this build’s runbook and state file.

Your own runbook

Answer six questions and copy or download a numbered runbook for your own move, in the order this build used.

Reportedn = 1 migration The order is this build’s, from its runbook and state file.

The migration kit, free

Measuredn = 1 kit Every script in the kit was run in dry-run mode on the PC this class describes before publishing; none was run with -Apply. MIT licence; read the README first.

How sure is each part of this page

MeasuredReportedCalculatedQuotedn = 1 migration Every figure on this page is an entry in the class’s evidence ledger, re-derived from the primary files on Nov 11, 2026.