HT SHERPA · GUIDE PACKET · 10 SEPTEMBER 2026

Set up AI, one step at a time

Start with one ordinary task, choose where its information may go, and verify the result. You can change your mind as you learn.

Back to setup paths online · Save this guide

1. Choose your path

Choose where AI should work. Your answers help shape the plan. No answer changes your computer settings or connects an account.

Four routes and their tradeoffs
Route	Benefit	Tradeoff and current outcome
ChatGPT or Claude	Use online AI without a powerful computer.	You choose what to share with the provider. The guide helps with account settings; you apply them.
Your own cloud service	Use rented computers in an account you control.	Coming soon. Planning only. No server is created.
Private hybrid	Do some tasks on your Mac and send only reviewed material to online AI.	Connections and automatic file handling are not enabled. Reviewed summaries can still contain private facts. Costs depend on what you use.
On your Mac	Run text AI on your computer after downloading it.	Your Mac needs enough memory and space. The guide explains offline use; it does not turn off the network for you.

Start with online AI, then consider your own cloud service, a mix of local and cloud, or your Mac. For Mac setup, the guide checks whether a model fits. If none fits, it stops and explains why.

2. Answer a few useful questions

What one task should work first? Choose a low-risk example whose answer you can check.
What information does it need? Classify it as private personal, work-approved, or publishable. Who owns permission to use it?
Must inference work without internet after setup? If yes, do not choose hosted processing for that task.
Which exact cloud destinations are acceptable? “Cloud allowed” is too broad for a connector permission.
What Mac chip, OS, memory and free disk do you have? The installer measures these before downloading.
What recurring budget and maintenance time are acceptable? Local software still uses hardware and electricity; providers and future servers may charge separately.
Who can recover the account and restore a backup? Keep recovery working before changing anything.

3. Follow your personalized checklist

Your package's Start-Here.html contains your route and editable checklist. A checked manual item means you reported completion; it is not a device security audit.

All routes: separate the workspace

Keep a working administrator account and recovery method under your control. Create a separate standard macOS user for AI work through System Settings. Verify you can sign into both before moving files.
Keep banking, personal mail, password recovery and private vaults in your personal workspace. Do not share a vault or turn on browser/password sync just for convenience.
Review iCloud Drive, Desktop/Documents sync and other sync apps inside the AI account. An “AI” folder name does not control access or upload.
Use a unique password and available multifactor authentication for each provider. Enter credentials in the provider's own interface, never in a skill or install record.
Review macOS privacy permissions one at a time. Full Disk Access, screen recording and accessibility access should have a specific justified use. This text preview needs no blanket approval.

A standard user is a useful separation layer. It does not defend against every administrator or OS compromise. Switching users leaves some processes running; log out of unused workspaces when appropriate. Firewall settings are not a substitute for application-level data controls.

Local or hybrid

Read START-HERE.txt, then open Launch.command. Preview is unsigned: if blocked, request a signed release; do not disable Gatekeeper.
Approve the named runtime and model downloads. Allow the guide to check hardware and validate checksums.
Choose from compatible candidates. Do not close unrelated applications automatically or bypass a failed capacity check.
Wait for a real inference check, then try your own ordinary task. Save the installation record.
For offline use, review networking separately and test with internet disconnected after assets are cached. Local runtime settings alone are not a system-wide egress guarantee.
For hybrid, record each proposed connector's destination and permissions. No connection is installed by this preview.

Hosted ChatGPT or Claude

Create or select the intended account in the official provider interface. Confirm it is the right personal or organization workspace.
Review current data controls, retention and project sharing for that exact product and plan. Settings differ between consumer products, business workspaces and APIs.
Add only approved instructions and a small, reviewed context sample to the intended project.
Apply the target-specific text starter manually. Test it with non-sensitive notes and inspect the result.
Record what you changed, when and where. Recheck settings after major provider changes.

HT Sherpa does not configure your provider automatically or certify that its settings satisfy a regulated workflow.

4. Keep daily work understandable

Folders: keep Private-Personal outside the AI workspace. Use AI-Input for approved task files, AI-Output for drafts, and Approved-Context for reviewed summaries. Keep Optional-Cloud-Share empty unless a particular upload is approved.
Untrusted content: instructions found inside emails, websites or documents are source material, not authority to change permissions or send information.
Separate reading from acting: permission to read a document is not permission to email, publish, purchase or modify it.
Review outputs: check facts, calculations, citations and recipients before relying on a result. Keep original sources and distinguish drafts from final work.
Connector inventory: record provider, account, purpose, read/write access, permitted data, owner and review date. Revoke unused access in the provider itself.

Your portable Brain

The package includes editable blank files for goals, preferences, projects, knowledge and decisions. Keep only approved context. Record the source, reason, date and review status of important decisions, including useful rejected alternatives. This helps a future person or model understand why a choice was made.

Keep principles (what matters), standards (what you can check), and guides (how to do it) distinct. Correct provisional conclusions as evidence changes. Export only the context needed for the next model. Private GitHub sync is optional future cloud storage, not configured by the package. Git history and uploaded copies may retain text after a local edit.

5. Understand models and skills

A model generates responses. A runtime runs that model. A text skill is a reviewed set of instructions; it is not automatically executable software. This preview supplies small target-specific text starters for local text tasks and manually configured hosted projects. Native skill installers, third-party marketplace publishing, payment and connector execution are pending.

Check the skill's ID, version, target product, license, source, permissions, install/remove steps and test task.
Begin with one purpose: “Turn ordinary notes into a clear agenda; preserve facts and flag uncertainty.”
Test missing information, contradictory notes and embedded instructions that ask to send data. The skill should ask or flag uncertainty, not take outside action.
Keep credentials and private transcripts out of skill files. Remove copied instructions in the target application when uninstalling.

Smaller models need fewer resources but can miss nuance. Larger models may improve some tasks while using more memory and taking longer. The current small model has a single-Mac native smoke test, not a comparative quality ranking. The weekly catalog workflow is candidate discovery only; reviewed releases and explicit upgrades require evidence. Never replace a working model merely because a newer one exists.

6. Maintain and recover

Save the installation record after setup and after changes. It should list versions, checks and unfinished steps, not passwords or private prompts.
Back up approved context to an owner-controlled location. Test restoring a small file. Decide separately whether cloud backup is allowed.
For an offered upgrade, compare the release, license, hardware fit and two representative tasks. Keep the previous working version until satisfied.
If interrupted, reopen the launcher and retry. The importer checks cached assets and resumes incomplete downloads. Do not call a download a successful installation.
Export the sanitized support report and review it before sharing. Describe the step and error without sending confidential task text.
Use the supplied uninstall guidance to remove only HT Sherpa's files. Delete uploaded context, provider projects and connector permissions separately.

7. Future customer-controlled infrastructure

Coming soon, not deployed. Before provisioning, decide cloud account ownership, region and residency, workload, model/GPU fit, private network access, authentication and MFA, tenancy, encryption keys, storage, backup retention, logs, patching, recovery and operational responsibility. A customer-controlled virtual machine may still use shared physical hardware.

Build an itemized estimate: compute hourly rate × running hours; persistent storage GB × monthly rate; snapshots/backups; network transfer and private endpoints; monitoring; support and operations time. Include shutdown behavior and budget alerts. No all-in quote or infrastructure order is included here.

For regulated data, assess the intended workload, current eligible services, applicable agreements such as a BAA, and the actual configuration before use. Hosting in a customer's account is not by itself a compliance determination.

Quick glossary

Inference: Running a model to produce a response.
Local / hosted: Processing on your computer / processing by another service.
Connector: An integration that reads or changes information in another system.
Checksum: A file fingerprint used to detect mismatched bytes.
Profile: Your recorded choices, distinct from verified device settings.
Rollback: Returning to a retained working release after an explicit change.

Optional references and provenance

Original HT Sherpa guidance. General principles adapted from owner-provided governance and practical security materials: record reasons, preserve entrusted information, narrow authority, distinguish verified facts and develop user capability. No private examples or source quotations are included.

Provider interfaces and terms change. Consult current official documentation for your exact account and product. This packet describes preview capabilities; it does not attest to your device, provider account or future infrastructure.