Turn your agent into a careful computer repair engineer
Evidence before conclusions; a plan before any change.
A cross-platform computer diagnosis and repair Skill for Codex, Claude Code, OpenClaw and other agents. Application and data migration, a C: drive turning red, a sluggish PC, unwanted software, recurring pop-up ads, network trouble or OpenClaw configuration issues — describe the symptom in one sentence. The agent gathers evidence, proposes a plan and changes nothing until you confirm; updates, printers, boot repair, recovery, BitLocker/PE checks and other workflows remain covered too.
4/7Propose a repair planNothing runs until you confirm
What it is, what it is not
Computer Repair Skill is not a tune-up utility. It is a computer-repair process that guides your agent to diagnose step by step, explain the risk, and fix the problem only after you confirm.
No background service and no takeover of your computer
How Computer Repair Skill positions itself
How Computer Repair Skill positions itself
No.
It is
It is not
1
A Skill package an agent readsIt supplies the diagnostic steps, confirmation rules and safety boundaries; your agent does the executing.
A desktop applicationNo installer, no user interface, no background process.
2
Files you copy into a Skills directory~/.codex/skills/, ~/.claude/skills/, ~/.agents/skills/.
A companion program, model or cloud APIThe Skill itself is Markdown and YAML — there is nothing extra to install.
3
Review the diff, the risk and the rollback path firstOnly after you confirm does it apply the smallest change that fixes the problem.
“One-click Windows optimization”Tuning presets can affect updates, privacy, drivers, enterprise policy and your ability to roll back.
4
64 on-demand playbooksLoaded through a routing index according to the problem, so irrelevant context stays out.
One giant prompt that fills the context windowIt loads only the process relevant to the problem, keeping unrelated steps out.
5
An auditable, reversible repair procedureWhenever state changes, the plan, the verification steps and the rollback command are kept.
A repair licence or a data-recovery guaranteeSmoke, liquid ingress, unusual noise, overheating, a swollen battery or a physically failing source disk means stop immediately and escalate to a professional.
How the agent troubleshoots, step by step
From describing the problem to verifying the result, the agent moves through seven clear steps. Before any change, it explains the reason, impact and rollback path, then waits for your confirmation.
1
Clarify the goal
Break the request into symptoms, the desired end state and the constraints that must not change. Multiple goals are kept item by item — no deleting instead of migrating, no reinstalling instead of repairing, no skipping an explicit requirement.
Action tiers: read-only audit · reversible change · high-impact change · system-level red line
2
Route to a playbook
Read the routing index and pick the single most specific playbook by symptom, platform and trigger phrase. Only the one the current task needs is loaded — never all of them at once.
A playbook is a diagnostic protocol, not permission to execute
3
Read-only diagnosis
Start with narrow checks that change nothing, recording commands, exit codes, measured values, paths, service states and timestamps. Command output is evidence, and a single anomaly is not a root cause.
Metadata only by default: no file contents, password stores, cookies or private keys
4
Propose a repair planmust stop
Before any install, upgrade, long-running task, administrator operation or state change, it first shows: the confirmed symptoms and most likely cause, the specific actions and their blast radius, whether elevation / network / reboot is needed, the rollback point, and the verification commands and success criteria for afterwards.
Wait for the user's explicit confirmation. Confirmation given earlier for a different set of actions does not cover a newly surfaced high-impact action.
the pivot of the whole Skill is here
5
Apply the minimal change
Only actions already shown get executed; the moment the scope widens, the plan is presented again. Literal paths and structured parameters are used instead of relying on the current directory or broad wildcards. Cleanups record, item by item, the target path, size/hash, original location, time and recovery location.
Prefer the recycle bin or a quarantine directory — never permanent deletion as a stand-in for “cleanup”
6
Verify the result
Re-run the diagnostics tied directly to the fault; a command exiting successfully is not the same as the problem being solved. Compare the data before and after, and check that adjacent functionality has not regressed.
Where a playbook defines success criteria, each one is met
7
Wrap up
Report concisely in your current language: the root cause or the most credible current conclusion, the commands actually run and what changed, the verification evidence and final state, plus any remaining risk and the rollback command.
Never claim a fix that has not been verified
Is it safe to let an agent repair your computer?
Deleting files, using administrator privileges, or changing disk and boot settings can carry serious risk. Computer Repair Skill grades every action by impact: the higher the risk, the more precise the scope, rollback plan and confirmation must be.
The user's confirmation for an ordinary repair does not extend to these objects.
Four risk tiers
L1
Read-only
Observes the system without changing state — querying the network, disks, processes, services, logs and file metadata.
May run directly once the scope is bounded
L2
Reversible change
Low impact with a clear rollback — flushing DNS, restarting an ordinary service, moving a single file.
Show the plan, impact and rollback, then get confirmation
Show the exact command, target scope, recovery plan and verification; confirm separately
L4
System-level red line
The boot volume, system root, partition table, firmware, BIOS/UEFI, system keychain, credential database, an entire user directory — or disabling a core security mechanism.
Stop execution and hand over to the vendor or enterprise recovery path
Six steps that must be completed before any deletion
Inspect Before Delete · applies to every deletion
Identify the target by literal absolute path, without relying on the current directory.
Use read-only commands to list the target's path, size, modification time, type and key contents.
Confirm that the inspection result really corresponds to the specific objects about to be deleted.
Show the target list, the reasoning, the expected impact and the recovery method.
Obtain the user's confirmation for these specific targets.
Delete the explicitly named items, then verify the state of the target and of neighbouring directories.
Never executed
These forms never reach the execution stage at all
Pipelines that hand a remote response straight to a shell: irm | iex, iwr | iex, curl | powershell, curl | bash, wget | sh, bash -c "$(curl …)".
Using *, recursing from a root, or sweeping an ancestor directory across user data, application data or credential directories.
Hiding deletion targets behind a pipe, xargs, eval, command substitution, an encoded payload or a dynamic variable.
Re-running the same deletion through a scripting runtime such as Python or Node to get around the host's guardrail on rm -rf or Remove-Item.
Printing passwords, recovery keys, private keys, tokens or full sensitive paths into the chat.
Actively disabling Defender, SmartScreen, UAC, the firewall, Windows Update or Core Isolation — these are audited for state only.
time | target | action | risk level | user confirmation | exit status | verification | rollback
Secrets are stripped when the command template is recorded; where a path contains a username or company information, only the part needed for diagnosis is kept.
How to install it in Codex, Claude Code or OpenClaw
Four ways, from “hand the agent one sentence” to “copy the directory yourself”. A client that only supports chat, with no local file or terminal access, cannot actually install it for you.
Hand the agent one sentence
Best for newcomers. Open Codex, Claude Code, Hermes, OpenClaw, Cursor or any other tool that supports Agent Skills, and send it the text on the right as-is.
The agent needs to be able to reach GitHub, read local files and run the necessary installation commands.
prompt
Please install this Agent Skill: https://github.com/88lin/computer-repair-skill
Review the repository contents and SKILL.md first, then install it into my current Agent's Skills directory; if an older version is already installed, back it up before updating.
When you are done, tell me the actual install path, the target Agent and the verification result.
Skills CLI
Fastest if Node.js and npm are already installed. The first run asks you to choose an install scope and a target agent; afterwards, restart the agent or refresh its Skills list.
Install scopes, target agents and link sources follow the Skills CLI's current documentation: skills.sh/docs.
bash
# Interactive install
npx skills add 88lin/computer-repair-skill
# Global install
npx skills add 88lin/computer-repair-skill --global
# Codex only / Claude Code only
npx skills add 88lin/computer-repair-skill --global --agent codex
npx skills add 88lin/computer-repair-skill --global --agent claude-code
Repository installer
For users who want explicit control over the path. codex, claude and agents use the corresponding default Skills root; custom together with -Destination / --destination puts it wherever you say.
When the target directory already exists the installer refuses to overwrite by default. To confirm an upgrade, pass -Force or --force explicitly; the old version is backed up first to an adjacent external/computer-repair-skill/backups.
powershell / bash
git clone https://github.com/88lin/computer-repair-skill.git
cd computer-repair-skill
# Windows PowerShell
.\scripts\install.ps1 -Target codex
.\scripts\install.ps1 -Target custom -Destination "D:\path\to\skills"
# macOS / Linux
./scripts/install.sh --target claude
./scripts/install.sh --target custom --destination "/path/to/skills"
Manual copy
For users who cannot use a command line. Copy the whole skills/computer-repair-skill directory out of the repository into your target agent's Skills root.
The directory must keep SKILL.md, agents/ and references/ — do not copy just one playbook.
From a C: drive turning red, a sluggish PC, unwanted software, recurring pop-up ads, network trouble, app migration, uninstall residue and large-folder management to updates, printers, boot repair, recovery, security, developer setup and OpenClaw. Search by symptom, or open any row to see when it applies and what you can ask.
By category
Category
Count
By platform
“Cross-platform” means the procedure does not depend on a single operating system — DNS, mail, identity services, SSH, Wi-Fi, Outlook, OpenClaw, credential and local-data audits.
JavaScript is disabled, so search, filtering and detail dialogs are unavailable. You can still browse all 64 playbooks in the table below.
Complete routing index for all 64 playbooks
No.
Playbook
Category
Platform
Route ID
01
🩺Comprehensive device health check — disk, memory, uptime, updates, firewall, backup, and network
Health, performance, storage & backup
Cross-platform
health-baseline-check
02
⚡Diagnose slowness, high CPU, memory pressure, and hangs
Health, performance, storage & backup
macOS
performance-forensics
03
🩺Diagnose Windows slowness, CPU saturation, memory pressure, disk bottlenecks, startup load, and application hangs
Health, performance, storage & backup
Windows
windows-performance-forensics
04
🩺Diagnose Linux load, CPU, memory pressure, swapping, IO waits, disk bottlenecks, cgroup limits, and process hangs
Health, performance, storage & backup
Linux
linux-performance-forensics
05
🧹Safe maintenance sweep for a generally-sluggish Mac — refresh caches and reclaim memory without risk
Health, performance, storage & backup
macOS
mac-tune-up
06
💾Reclaim disk space on macOS — measure, audit, clean cache-class data, report freed bytes
Health, performance, storage & backup
macOS
disk-space-recovery
07
🩺Reclaim Windows disk space by measuring volumes, preserving user data, and verifying freed space
Health, performance, storage & backup
Windows
windows-disk-space-recovery
08
🩺Build a privacy-preserving Windows storage inventory that explains large paths before any cleanup
Health, performance, storage & backup
Windows
windows-storage-inventory
09
🩺Discover, lazily size, and safely offload selected Windows system or application folders
Health, performance, storage & backup
Windows
windows-large-folder-management
10
🩺Migrate an application or selected data folder to local NTFS storage with junction rollback
Health, performance, storage & backup
Windows
windows-application-migration
11
🩺Audit Windows migration Junctions, detect ghost records, and restore data safely
Health, performance, storage & backup
Windows
windows-migration-history-recovery
12
🩺Audit and clean regenerable Windows app caches or duplicates while preserving profiles, chat data, and credentials
Health, performance, storage & backup
Windows
windows-application-cleanup
13
🩺Audit and clean regenerable macOS app caches or duplicates while preserving containers, chat data, and keychains
Health, performance, storage & backup
macOS
macos-application-cleanup
14
🩺Audit and clean regenerable Linux app caches or duplicates across XDG, Flatpak, and Snap data roots
Health, performance, storage & backup
Linux
linux-application-cleanup
15
🩺Audit Windows app install/uninstall candidates by exact package ID, publisher, source, and reversible verification
Health, performance, storage & backup
Windows
windows-application-lifecycle-audit
16
🩺Inspect a new Windows computer, preserve unboxing evidence, verify support/warranty, and establish a safe baseline
Health, performance, storage & backup
Windows
windows-new-device-intake
17
🩺Reclaim Linux disk space and inodes by measuring filesystems, cleaning native caches, and verifying services
Health, performance, storage & backup
Linux
linux-disk-space-recovery
18
📂Audit a device for locally stored sensitive or company data — for offboarding or device reassignment
Health, performance, storage & backup
Cross-platform
local-data-audit
19
💾Verify backup integrity by checking status, timestamps, and testing a real file restore
Health, performance, storage & backup
Cross-platform
backup-verify-restore
20
🛡️Set up Time Machine backup on macOS or configure basic backup strategy
Health, performance, storage & backup
macOS
setup-backup
21
🌐Systematic connectivity troubleshooting for Wi-Fi, DNS, and internet issues
Network, DNS, VPN, identity & mail
macOS
network-diagnostics
22
🌐Systematic Windows connectivity troubleshooting for adapters, DHCP, routes, DNS, proxy, VPN, and HTTP
Network, DNS, VPN, identity & mail
Windows
windows-network-diagnostics
23
🌐Systematic Linux connectivity troubleshooting for links, addresses, routes, DNS, HTTP, firewall, and namespaces
Network, DNS, VPN, identity & mail
Linux
linux-network-diagnostics
24
🔒Diagnose VPN connection failures, drops, and split-tunnel DNS issues
Network, DNS, VPN, identity & mail
macOS
vpn-troubleshooting
25
📋Check email-related DNS records (MX, SPF, DKIM, DMARC) for a domain
Network, DNS, VPN, identity & mail
Cross-platform
dns-record-check
26
📡Test network connectivity to email servers — SMTP, IMAP, POP3, and major providers
Network, DNS, VPN, identity & mail
Cross-platform
email-connectivity-test
27
🪪Test connectivity to identity providers — Microsoft Entra ID, Google Workspace, and SSO endpoints
Network, DNS, VPN, identity & mail
Cross-platform
identity-provider-test
28
📶Connect to a new Wi-Fi network including enterprise/WPA2-Enterprise networks
Network, DNS, VPN, identity & mail
Cross-platform
setup-wifi-profile
29
✉️Add an email account to Apple Mail or Outlook on macOS
Network, DNS, VPN, identity & mail
macOS
setup-email-account
30
🩺Fix app crashes, launch failures, and permission issues
Apps, updates & printing
macOS
app-doctor
31
📧Fix Outlook sync failures, crashes, stuck email, and profile corruption
Apps, updates & printing
Cross-platform
outlook-troubleshooting
32
🔄Fix stuck macOS updates, failed installations, and software update errors
Apps, updates & printing
macOS
update-troubleshooting
33
🔄Fix stuck Windows Updates, failed installations, and update service errors
Apps, updates & printing
Windows
windows-update-troubleshooting
34
🖨️Fix stuck print jobs, missing printers, and CUPS issues
Apps, updates & printing
macOS
printer-repair
35
🖨️Fix stuck print jobs, offline printers, and spooler issues on Windows
Apps, updates & printing
Windows
windows-printer-repair
36
🧩Review reversible Windows privacy, AI, performance, and usability settings with version gates, diff, and rollback
Apps, updates & printing
Windows
windows-configuration-review
37
🩺Preview and clean Windows uninstall residue with strict identity matching and recovery
Apps, updates & printing
Windows
windows-uninstall-residue-cleanup
38
🪟Diagnose Windows driver failures with hardware IDs, vendor source, reinstall, rollback, and compatibility checks
Windows repair, boot & hardware
Windows
windows-driver-lifecycle-audit
39
🪟Troubleshoot Windows camera, microphone, headset, and speaker failures across hardware, permissions, drivers, and apps
Windows repair, boot & hardware
Windows
windows-av-input-triage
40
🪟Triage Windows startup, No Boot Device, UEFI/boot-entry, and pre-login failures without blindly changing firmware
Windows repair, boot & hardware
Windows
windows-boot-failure-triage
41
🪟Diagnose Windows startup/system corruption with WinRE, Safe Mode, SFC, DISM, and restore paths
Windows repair, boot & hardware
Windows
windows-winre-system-repair
42
🪟Audit BitLocker/device encryption before PE, BIOS, hardware, partition, repair, or reinstall operations
Windows repair, boot & hardware
Windows
windows-bitlocker-recovery-triage
43
🪟Audit Windows partition and C-drive expansion requests with backup, BitLocker, adjacency and offline-tool safety checks
Windows repair, boot & hardware
Windows
windows-partition-resize-audit
44
🪟Plan safe Windows laptop hardware maintenance for parts, battery, dust, and thermal work without automatic execution
No playbook matches. Try another keyword, or clear the filters.
How to describe your problem to the agent
After installing, describe the problem directly — there is no playbook name to memorise. Include the platform + symptom + constraint + expected action so the agent can choose the right troubleshooting process.
01
My C: drive is turning red. First inventory which categories use the most space, then give me a risk-tiered cleanup plan — do not delete anything without my approval.
This Windows PC is getting slower and takes ages to boot. First identify the processes using the most CPU, memory and disk, then list the likely causes; make no changes until you propose a fix.
Unwanted software keeps spawning pop-ups and starting with Windows; it may have left residue after uninstalling. First list startup items, scheduled tasks and other evidence read-only. I will confirm each cleanup item; explain how to restore it.
WindowsRelated playbookwindows-persistence-audit
04
My system disk is running out of space, and I want to move some apps to D:. Check available space, file locks and the rollback plan first; verify the copy before switching the original path, and keep it reversible.
Wi-Fi is connected, but webpages will not load. Trace the path layer by layer, from DNS and proxy through routing; do not reset the network or change browser settings without my approval.
WeChat and Chrome are using tens of GB on my Mac. First separate the regenerable caches from session state and chat history, read-only; once I confirm, clean only the caches.
macOSRelated playbookmacos-application-cleanup
08
A Flatpak app's data directory on this Linux box is huge. Confirm its packaging format and data root first, separate cache from my own data, then give me a reversible cleanup plan.
LinuxRelated playbooklinux-application-cleanup
Standard order of work: confirm platform and constraints → gather evidence → explain candidate causes → propose the minimal change → wait for confirmation → execute → verify and record the result.
Frequently asked
Straight answers about installation, agent capabilities, safety boundaries and the open-source licence.
01
A Skill, not a background appIt supplies the workflow; execution still depends on the tools and permissions available to your agent.
02
Read-only first, confirm before changesInstalls, deletion, elevation and restarts must show their impact and rollback path before they run.
03
Describe the problem, not the playbookInclude the platform, symptom, constraints and desired action so the agent can route accurately.
No. It is a Skill package an agent reads, supplying diagnostic steps, confirmation rules and safety boundaries. Whether anything actually runs depends on whether your agent has a local terminal, file access, network access and a confirmation mechanism.
You can have the agent explain the procedure and generate a checklist and manual steps; but without local tools it cannot read your real system for you, and it must not claim anything has been fixed or verified.
Because tuning presets can affect updates, privacy, drivers, enterprise policy and your ability to roll back. Computer Repair Skill reviews the diff, the risk and the rollback path first, then applies the minimal change once you have confirmed.
Do not assume so. Before entering PE/WinRE, have the agent save the BitLocker status, backup checks, command list and stop conditions; a human works through the list during the offline phase, and the agent reviews the result once you are back in a normal system.
Codex, Claude Code, Hermes, OpenClaw, Cursor or any other tool that supports Agent Skills. The installer has three default Skills roots built in — codex, claude and agents — plus custom for a path you specify.
No. It starts with read-only checks and explains what it finds. For installs, deletion, elevated access, restarts or any other state change, it first shows the exact action, scope, impact and rollback path, then waits for your explicit confirmation.
The project is open source under GNU AGPL-3.0. You can inspect, review and modify the source; when distributing a modified version, follow the same licence's source-disclosure and attribution requirements.
When your computer acts up, let your agent find the cause first.
Install it with one command. When something goes wrong, tell your agent the platform, the symptom and your constraints; it will gather evidence first, then propose the next step.