# Skill: Explain why local AI belongs on a user's own machine and when it beats the cloud

> Apply this skill when a user asks whether to run AI locally or stick with cloud services, or when you need to argue the trade-offs of on-device inference.

Source: sodigi-learn -- local-agents/zachem-lokalnyy-ii -- https://sodigi.io/learn/local-agents/zachem-lokalnyy-ii

## When to use
- User asks whether their data is safe when talking to AI services.
- User wonders if they can work with AI offline (travel, plane, cabin, poor connectivity).
- User wants to avoid unpredictable token bills or cloud usage limits.
- User is evaluating local vs cloud for a new project and needs an honest decision framework.

## Core rules
- Local AI means the model runs on the user's own hardware; no data leaves the machine at any point.
- Privacy is the primary argument: code under NDA, client documents, health/financial data must never be uploaded to a third-party server.
- Cost after install is electricity only -- the open model weights are free to download and free to run.
- Offline capability requires the model to be downloaded in advance; it cannot be fetched mid-flight.
- Local models are weaker than top cloud models on hard reasoning tasks -- do not oversell them.
- The bottleneck is memory (RAM and especially VRAM), not CPU speed.
- The recommended default: local for sensitive and routine work, cloud for rare hard tasks. Never forced to pick only one.
- Never claim local AI is "completely free" -- electricity and hardware wear are real costs, just predictable.

## Procedure
1. Identify the data sensitivity of the task: can this data legally or practically leave the user's machine?
2. Assess task complexity: everyday writing, summaries, simple code, drafts -- local handles these well. Deep reasoning, long chains, novel research -- cloud still leads.
3. Assess the user's hardware: ask about available RAM and GPU VRAM before recommending a model size.
4. If data is sensitive AND hardware is adequate: recommend local unconditionally.
5. If data is non-sensitive AND task is hard: recommend cloud without hesitation.
6. If both or neither: recommend a hybrid -- local for daily work, cloud only for the heavy outliers.
7. Remind the user to download the model before going offline; the download itself requires internet.

## Ready-to-use prompt
```
I want to help the user decide whether to run AI locally or in the cloud.

Context to gather first:
- What kind of data will they feed the AI? (code, personal docs, public content, etc.)
- Do they work offline or travel frequently?
- What are their hardware specs? (RAM, GPU VRAM if any)
- How often do they need AI, and do they hit cloud quotas?

Decision logic:
1. Sensitive data (NDA code, client files, health, finance) -> local, no debate.
2. Hard tasks (complex reasoning, long analysis) with non-sensitive data -> cloud is better.
3. Routine tasks (summaries, drafts, simple Q&A) -> local saves money.
4. Mixed workload -> hybrid: local as default, cloud only for the hard stuff.

Key trade-offs to communicate honestly:
- Local is private, free after install, works offline, no usage limits.
- Local models are weaker on hard tasks and need decent hardware (memory is the bottleneck).
- Cloud is smarter and always available but paid, dependent on internet, and your data leaves your machine.
- A hybrid strategy almost always beats choosing one side only.
```

## Pitfalls
- Claiming local AI is worse at everything: for most everyday tasks the quality gap is negligible.
- Recommending a model before asking about available VRAM; a model that doesn't fit RAM will crawl or crash.
- Telling the user they'll have offline access without reminding them to download the model first.
- Portraying local as "free" without mentioning electricity and hardware wear.
- Forcing a binary choice when a hybrid is almost always the practical winner.
- Recommending the biggest available model "because it's smarter" without checking if the hardware can run it.